chore(deps): raise the sqlalchemy floor on 3.14; drop the GIL workaround

[lesnik512]

Closes #160.

Why now

#160 waits on SQLAlchemy's cyextensions declaring Py_MOD_GIL_NOT_USED. They do
as of 2.0.54. Measured on CPython 3.14.7t with the compiled extensions
genuinely loaded (has_compiled_ext() == True, collections.cpython-314t-*.so),
no env var set:

sqlalchemy	compiled ext	GIL after import
2.0.52 (the version in #160)	yes	enabled
2.0.53	yes	enabled
2.0.54	yes	disabled

So DISABLE_SQLALCHEMY_CEXT_RUNTIME=1 now buys a slower pure-Python fallback and
nothing else.

What

• Floor: sqlalchemy[asyncio]>=2.0.54; python_version >= '3.14'. The existing
  marked floor is split so ==3.13 keeps 2.0.31 and below 3.13 keeps 2.0.
• CI: the env: block goes from the freethreaded job. The job already
  asserts sys._is_gil_enabled() is false, so the floor is now self-gating:
  a regression turns that step red instead of being masked by the workaround.
• Docs: the admonition in introduction/installation.md kept its second half,
  which is still true (a foreign-broker client like aiokafka still re-enables
  the GIL), and lost the SQLAlchemy instruction.
• ADR-0002 recorded the workaround as load-bearing. Amended rather than left
  standing, since docs/agents/domain.md tells agents to read ADRs as current
  truth about the area they are working in. The decision itself is unchanged;
  only its bounding constraint moved.

The cost, stated plainly

There is no environment marker for a free-threaded build, so python_version >= '3.14' is the narrowest expressible constraint and ordinary 3.14 users inherit a
2.0.54 floor for a reason that only affects 3.14t. That is the trade for the docs
being able to state the guarantee without a conditional the reader has to
evaluate. Option 2 in the alternative was to keep the floor and rewrite the note
as "only needed below 2.0.54"; this is option 1.

Verified

• The freethreaded job's exact assertion on 3.14t with no env var: passes,
  GIL disabled with sqlalchemy 2.0.54 | compiled ext: True.
• Full suite on 3.14t without the workaround: 517 passed, 107 skipped.
• The lowest leg added in ci: resolve the declared dependency floors on every PR #188 resolves the new floor and passes:
  3.13 to sqlalchemy 2.0.31, 3.14 to 2.0.54, both 517 passed, 107 skipped.
  (The 107 are the Postgres tests, which skip locally and run in CI.)
• just lint-ci and just docs-build clean.

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	953	1.000	6.79	917	242	5000	5000	10000
consumer/w1/b100	951	1.000	6.83	920	243	5000	5000	10000
consumer/w2/b10	1100	1.000	6.80	961	242	5000	5000	10000
consumer/w2/b100	1295	1.000	6.78	992	263	5000	5000	10000
consumer/w4/b10	974	1.000	6.74	949	243	5000	5000	10000
consumer/w4/b100	1496	1.000	6.86	1032	243	5000	5000	10000
consumer/w1/b100/tfbs100	4967	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	2082	0.000	3.04	584	0	0	0	0

Gated (fails the build): delete_calls + tuple counters (upd/del/ins) + the producer's insert_calls, exact; select_calls within +2; wal_records within a 10% band. msg/s, WAL bytes and total calls are informational (timing/FPI noise).