perf!: make importing Auto3D 44x cheaper, and harden the tests that guard the coming refactors

[isayev]

Wave 1 of the remaining debt closure: the test hardening that has to land before the
refactors move the code it covers, plus the import-cost fix.

The headline

import Auto3D no longer imports torch or RDKit.

	before	after
import Auto3D	1.35 s	0.031 s
len(sys.modules)	1175	154 (stdlib floor is 128)
torch / RDKit loaded	yes	no
Auto3D.* submodules loaded	20	0

Three eager optional-dependency probes in __init__.py defeated the _LAZY_API
mechanism that exists to prevent exactly this. Probing for ANI2xt reached batch_opt,
which reached the utils barrel, which reached validation, which imported torch and
models.loading. Importing the package paid for the entire package before the caller
asked for anything.

The probes are deleted rather than lazified: they are not public API, and every real probe
already exists at its use site. A comment in the file records that and says not to
reintroduce one.

Breaking: Auto3D.ANI2xt, Auto3D.warnings, Auto3D.version and
Auto3D.PackageNotFoundError are gone from the package namespace. None was public API —
all four leaked in via those probes and module-level imports. Documented public names are
unaffected and still resolve lazily. Recorded in the CHANGELOG.

Why the test hardening comes first

A test that asserts nothing cannot tell you whether a refactor preserved behavior, and
hardening it after the refactor moved its code proves nothing about the move. So the
hardening is scoped by a rule rather than by judgment: harden a finding now if and only if
the test it names covers a src/ file that the next refactors touch. Everything else is
deferred to the per-module cleanup sweeps.

Re-verifying the vacuous-test corpus against current source changed the numbers
substantially. The "~105 findings" is 64 distinct finding-groups — that report's Tier 3
list has 24 entries but only 15 distinct, the other 9 being cross-references to items
already counted, and the ~105 counted test functions rather than findings. Of the 64,
26 are in scope for this PR and 38 are deferred. Unlike the dead-code findings, none
had decayed: 31 were re-verified individually and all 31 were still exactly as vacuous as
reported.

Two of the three tests the audit called missing already exist, and are more thorough than
the literal ask — the force-sign toy-NNP test and the end-to-end pipeline assertions. Only
the narrow gaps were filled.

Tests that were passing for the wrong reason

• test_embed_single_with_dynamic_conformers used ethane, which is conformationally
  degenerate and always collapses to one conformer, so the assertion held regardless of
  whether dynamic conformer counting worked. Now hexane, bounded against
  calculate_conformer_count.
• test_chunk_meta_structure built its expected dict by hand from the same literal it then
  compared against. Now derived from ChunkMeta's own annotations and required keys.
• test_fire_independent_molecule_tracking asserted that per-molecule state existed, not
  that it was independent. Rewritten as a two-phase protocol — bootstrap with every
  molecule progressing, then make exactly one oscillate — asserting the exact a and dt
  values that follow.
• test_one_bad_molecule_does_not_remove_the_others parsed IDs with a stale
  split("_")[0], which made a Tier 1 test vacuous.
• Three pytest.raises had no match=: one caught bare Exception, and two could have
  passed on any unrelated ModelLoadError rather than the argument-order check they exist
  for.

Coverage added where a path had none

• The arity check's len(required) > 3 half was never exercised — only too-few-arguments.
• The alias vocabulary is now parametrized over the full synonym set in correct and
  transposed order (14 cases), calling validate_custom_nnp directly so nothing pickles.
• forces = -grad is written separately in ANI2xtAdapter and ANI2xAdapter and was
  tested in neither. Both now assert -2*coords against a toy quadratic model.
• The "produced no conformers after clash relief" warning had no test at all. One now
  drives the real embedding path with relieve_clash stubbed by atom count, asserting the
  warning fires once, names the right molecule, and that a sibling species survives without
  being warned about.
• Padding invariance gains an ANI2x case at 1e-3, slow-marked. Not run locally — torchani
  is absent and no NNP may be loaded on that box, so CI verifies it.

Two defects pinned rather than fixed

Both are xfail(strict=True), so they cannot be forgotten and cannot be fixed silently.
Both belong to clusters landing later.

1. filter_unique_optimized raises KeyError on a record with no E_tot, while the
   legacy filter_unique tolerates it. The two conformer filters were given a shared
   duplicate criterion but still disagree on malformed input — and the stricter one is the
   survivor, so this must be settled before the other is deleted.
2. opt_steps carries two different minimums. FIELD_BOUNDS declares ("ge", 1)
   while validation.py hand-writes >= 10 in two places. The test asserts the two
   agree rather than picking a value, so whichever is authoritative can win.

Two tests deleted rather than hardened

test_config_schema_exists cannot fail once the module imports, and no field on
Auto3DOptions has a mutable default for test_immutable_default_list to guard.

Slow-tier correction

test_calc_spe_uses_model_factory and test_model_name2model_calculator_uses_factory were
slow-marked despite monkeypatching every model construction. Both move to the fast tier.
Each file's module-level pytestmark became per-test decorators so the 15 genuine real-NNP
tests in those two files stay slow — verified by collecting -m "not slow" and confirming
exactly those two appear.

Verification

• 1302 passed, 9 skipped, 2 xfailed (fixed order), re-run on the committed tree.
• Import-cost fix: five tripwires confirmed red first for their stated reasons, then each
  fix mutation-verified by reverting it and confirming the named test goes red — including
  reintroducing the ANI2xt probe (4 red) and adding a cache to __getattr__ (the
  non-caching test red).
• Cost tests assert module count, not wall-clock seconds, and run in a subprocess since
  the parent pytest process has already imported everything.
• __getattr__ deliberately does not cache. Caching turns it into a capturing binding,
  which is how a lazily imported reader holds a stub permanently while monkeypatch
  reports success; test_lazy_torchani_import documents that failure surfacing 182 tests
  downstream. A test asserts the non-caching property.
• Eight-seed order-independence sweep.

Follow-ups this PR does not do

• get_device, CustomNNP and IsomerEngineFactory are in api.rst but not __all__;
  generate_conformers is the reverse. Reconciling them needs an owner holding both
  api.rst and test_public_api.py.
• batch_opt/batchopt.py:17-28 has twin dead probes. Harmless now that nothing imports
  batchopt at package-import time, but still dead; goes to the cluster that owns that
  file.