fix(psycopg): check connection health on pool checkout

[RitiGrover]

Description

PsycopgClient.create_connection builds the psycopg_pool without a check callback, so getconn() never validates a connection before handing it out. A connection silently killed by the network (idle firewall/LB timeout, DB restart) sits in the pool looking fine until something tries to use it, then fails hard.

Motivation and Context

Fixes #2007. Users see psycopg.OperationalError: the connection is closed on in_transaction() after the app has been running a few days, with no automatic recovery.

Passing check=AsyncConnectionPool.check_connection uses psycopg_pool''s own built-in liveness probe on checkout, so a dead connection gets replaced transparently instead of surfacing to the caller.

How Has This Been Tested?

• Added a regression test asserting the pool is constructed with the check enabled
• Verified it fails against the old code and passes with the fix
• Ran the existing backend test suite (test_connection_params.py, test_password_factory.py, and the rest of the non-DB-dependent tests under tests/backends/) - all pass
• ruff and mypy clean on the changed files

Checklist

• My code follows the code style of this project.
• My change requires a change to the documentation.
• I have added the changelog accordingly.
• I have read the CONTRIBUTING document.
• I have added tests to cover my changes.
• All new and existing tests passed.

[codspeed]

Merging this PR will not alter performance

✅ 24 untouched benchmarks

---

_{Comparing RitiGrover:fix/psycopg-pool-connection-check (180ac4d) with develop (8477e47)}

[waketzheng]

1. Consider adding a behavioral test that simulates a dead connection and asserts getconn() replaces it, in addition to the construction-level test.
2. Add a brief note about the per-checkout probe overhead (even if negligible)
3. Confirm the CHANGELOG entry is present and references #2285.

[RitiGrover]

Addressed all three: added a behavioral test that kills a real backend connection and confirms the next ORM call still succeeds (confirmed it fails with psycopg.errors.AdminShutdown without the fix), added a note on the per-checkout probe overhead in the changelog, and pointed the changelog entry at this PR instead of the issue.