docs: document param-ordering, agg-validation, and Attr dataclass invariants

[Xof]

Comment-only pass

No behavior change. Documents three invariants that were load-bearing but
undocumented; the rest of the tree was left as-is (it is already densely and
accurately commented).

• compiler.py — ADD a third hard invariant to the module header: parameters
  are appended to self.params in the same left-to-right order their
  placeholders appear in the emitted SQL. Positional paramstyles (qmark ?,
  format %s) bind the Nth value to the Nth placeholder, so any multi-source
  compile method must visit sources in textual order; Division is the subtle
  case (outer dividend FROM → divisor → re-visited inner dividend).
• aggregates.py — EXTEND output_domain to record the cross-file invariant
  that Grouping.__post_init__ pre-validates aggregate attrs, so the
  schema[self.attr] lookup cannot raise in normal flow.
• predicates.py — EXTEND the Attr.__hash__ note: a hand-written
  __eq__/__hash__ coexists with @dataclass(frozen=True) because dataclass
  only fills in dunders not already defined; the relation nodes take the
  opposite eq=False route.

Not addressed here (flagged for a later decision)

tests/test_postgres.py comments describe psycopg as using the format
paramstyle (%s), but psycopg 3.x reports pyformat (%(pN)s + dict params) —
as engine.py's own corrected comment already notes. The tests pass; only the
comments misdescribe the mechanism. Left untouched (comment/code mismatch to
correct deliberately, not paper over).

Local gate green: ruff, mypy, pytest (119 passed, 15 skipped).