Record FedAuth token deterministically in mock server

[saurabh500]

Summary

Fixes two design flaws in mssql-mock-tds that made downstream FedAuth regression tests (microsoft/mssql-python#652: est_unique_access_token_transmitted_exactly, est_distinct_tokens_on_sequential_connects) flaky/failing only on Linux.

Stacked on #80 — base branch is saurabh500-publish-python-pkg-ado-feed, not main.

The two flaws

1. Token recorded only at connection teardown, asynchronously. Each connection is handled in a spawned task; the token was captured into the per-connection ConnectionProcessor during login but only copied into the shared, queryable ConnectionStore after the read loop broke on client EOF. That insert ran on the server task after the client's close() had already returned, so tests had no barrier to wait on and papered over it with time.sleep.
2. Store keyed by client socket address. ConnectionStore was HashMap<SocketAddr, ConnectionInfo>. Two sequential connects from the same client process frequently reuse the same ephemeral local port on Linux, so the second connection's insert overwrote the first under the same key — has_received_token(first) then returned False.

The fix

• Record eagerly, before LoginAck. ConnectionProcessor now holds an Option<Arc<Mutex<ConnectionStore>>> and upserts its state into the shared store in both FedAuth paths (inline-token Login7 and challenged FedAuthToken) before the response bytes are returned. Because the client blocks on the LoginAck read, the token is guaranteed visible the moment connect() returns — no teardown dependency, no sleeps. The existing teardown store.store() calls remain as a harmless final upsert.
• Re-key by a unique per-connection id. An AtomicU64 counter on the server assigns a conn_id at accept() time, threaded through into the ConnectionProcessor. ConnectionStore now uses BTreeMap<u64, ConnectionInfo> so iteration and .values().last() are ordered and get_last_access_token() deterministically returns the most recent connection. store() upserts by conn_id (eager + teardown records for one connection update the same entry). ConnectionInfo.addr is retained for info; ConnectionStore::get now takes a conn_id.
• Removed the time.sleep workarounds in mssql-py-core/tests/rs-only-tests/test_mock_server_fedauth.py; conn.close() calls kept.

Validation

• cargo bclippy — green (-D warnings).
• rustfmt --check on the changed file — green.
• cargo nextest for mssql-mock-tds unit tests (16/16), mssql-tds test_mock_server_fedauth (all pass), test_redirection (all pass), and test_mock_server user-agent interception (pass). These Rust integration tests exercise the exact new eager-record + BTreeMap store code over TLS on Windows.
• Both Python bindings (mssql-mock-tds-py, mssql-py-core) build via maturin develop and import cleanly.

Note: The in-repo Python fedauth tests could not be run to green on Windows due to a pre-existing, Windows-only native-tls/schannel handshake limitation with the self-signed test cert (connect times out during the TLS handshake, before any login/token logic). This is orthogonal to this change; the downstream tests run on Linux, and the Rust integration tests validate the same code paths.

[saurabh500]

Superseded by #91, which targets main directly and contains only the mock-server token-capture changes (this branch was stacked on #80 and pulled in unrelated commits when retargeted).