Stage 1: mssql-tds reset correctness (ADO #47337)

[Saurabh Singh (saurabh500)]

Summary

Stage 1 of the mssql-odbc connection pooling effort (ADO User Story #47317). This is the bottom implementation layer, scoped to mssql-tds only — it makes a physical connection reset actually clean the session-bound client state so a pool can safely hand a reused connection to the next borrower. mssql-odbc wiring comes in later stages.

Implements Workstream A (A1, A3) and discovery D3 from docs/odbc-connection-pooling-plan.md.

A1 — Centralize the ResetConnection ENVCHANGE acknowledgement

Previously the ResetConnection ENVCHANGE was handled in 5 token-draining spots that each only called session_state_table.reset(), leaving session-bound client caches stale after a real server reset.

• New private TdsClient::on_reset_connection_ack() performs the full post-reset transition: resets the session-state table, clears prepared_handles / prepared_param_encryption / pending_capture / pending_prepared_param_encryption, and restores the negotiated DATABASE / LANGUAGE / COLLATION back to their login values (so a borrower's USE otherdb / SET LANGUAGE doesn't persist across a reset).
• The cache-clearing is factored into a shared clear_session_bound_caches() helper reused by the reconnect path.
• All 5 ResetConnection sites now call the single method; the following capture_change_property ordering is preserved (login-default restore moved out of execution_context into the single owner to avoid duplication).
• Per D5 (verified), no post-login ANSI SET-option batch is re-emitted — ANSI defaults ride the fODBC LOGIN7 bit that sp_reset_connection restores. No BuildServerSideConnectOptions equivalent added.

D3 — Mark the transport dead on fatal server error tokens

known_dead was only set on socket write/read failure, EOF, or explicit close; a class ≥ 20 fatal error on a still-open socket left is_connection_dead() reporting alive.

• Added TdsTransport::mark_known_dead() (default no-op; overridden by NetworkTransport).
• The shared error-token path (record_error_token) now marks the transport known-dead when severity >= 20, so a later pool checkout discards it.

A3 — Tests

• Unit test asserting the reset transition clears all four caches and restores DATABASE/LANGUAGE/COLLATION.
• Unit tests for the D3 fatal-error path (severity ≥ 20 marks dead; severity < 20 does not).
• Live test in tests/test_reset_connection.rs proving a prepared handle is invalidated across a reset (gated on a live SQL Server via .env, like the sibling reset tests).

Validation

• cargo bfmt ✅
• cargo bclippy (warnings-as-errors) ✅
• cargo nextest — new/related tests pass. Pre-existing certificate/TLS-validation test failures are unrelated to this change (they fail on the base branch too, expired cert fixtures).

Notes

• Stacked on base PR Add ODBC connection pooling design plan (ADO #47317) #293 (saurabh500-odbc-connection-pooling-plan); this PR targets that branch, not main.
• Refs [AB#47337](https://sqlclientdrivers.visualstudio.com/b95cf060-8083-439d-8ef1-405d5bf219d8/_workitems/edit/47337).