Stage 3: reset acking + isolation carrier (ADO #47339)

[Saurabh Singh (saurabh500)]

Summary

Stage 3 of the mssql-odbc connection pooling effort (ADO User Story #47317, Task [AB#47339](https://sqlclientdrivers.visualstudio.com/b95cf060-8083-439d-8ef1-405d5bf219d8/_workitems/edit/47339)). Makes the connection reset self-acking at pool checkout so a borrower never inherits a session whose reset has not yet been processed, and confirms the isolation/autocommit checkout lifecycle carries and survives the reset.

Stacked on Stage 2 PR #296 (→ #294 → #293). Base branch: saurabh500-stage-2-odbc-pooling-primitives, not main.

A2 — reset is self-acking (mssql-tds)

• New TdsClient::reset_connection(): arms the full RESETCONNECTION bit and drives a minimal SELECT 1 round trip so the server processes the reset and returns its ResetConnection ENVCHANGE within the call. The ack flows through Stage 1's on_reset_connection_ack(), clearing session-bound caches before checkout. A failed round trip returns an error so the caller can discard the connection.
• prepare_reset_connection(bool) is unchanged (the deferred-piggyback path); reset_connection() is the eager variant.
• New TdsClient::mark_connection_dead() lets the ODBC layer deterministically poison a client whose reset failed.
• Unit test proves the round trip carries the RESETCONNECTION bit and the ack clears session-bound caches.

B4 — wire eager reset + isolation carrier (mssql-odbc)

• SQLSetConnectAttr(SQL_ATTR_RESET_CONNECTION) now calls reset_connection() after arming, so the reset is processed+acked before the set-attr returns. The DBC mutex is never held across the round-trip I/O (claim → drop lock → I/O → restore). On failure it poisons the client and surfaces 08S01 so mssql-python discards it at checkout.
• Confirmed the SQL_ATTR_TXN_ISOLATION handler still emits a real SET TRANSACTION ISOLATION LEVEL batch; because the reset now acks eagerly, a later short-circuited isolation SET can't leave the reset unacknowledged. Test covers a borrower raising isolation to SERIALIZABLE, resetting, then re-applying READ COMMITTED at checkout.
• Documented the D9 raw-T-SQL isolation caveat (stale cache → possible leak; shared, documented limitation, not fixed here) as a code comment at the same-value short-circuit.

B5 — autocommit / EndTran lifecycle confirmation

• Added a checkout-cycle test: acquire (reset, eagerly acked) → setautocommit(False) (begin-txn) → check-in → next acquire (reset) reuses the same physical connection and leaves it reusable. No autocommit/EndTran plumbing rebuilt.

Tests

• New/updated unit + mock tests in mssql-tds and mssql-odbc: eager-ack round trip, failure poison → 08S01, reset+isolation re-apply, and the full checkout lifecycle. Extended test_client_support with a ResetConnection ENVCHANGE token helper and mark_known_dead support on the replay transport.

Validation

• cargo bfmt ✅ (workspace + mssql-py-core)
• cargo bclippy ✅ (warnings-as-errors, clean)
• mssql-tds / mssql-odbc lib unit tests ✅ — the only failures are the pre-existing cert/TLS-fixture tests on the base, unrelated to this change.

References [AB#47339](https://sqlclientdrivers.visualstudio.com/b95cf060-8083-439d-8ef1-405d5bf219d8/_workitems/edit/47339).

[Saurabh Singh (saurabh500)]

Consolidated into #308, which contains all of this work in a single PR against main. Closing in favor of #308.