fix: six small issues found during the sync/async parity review

[AhmadMasry]

Description

While preparing the async (asyncio) phase of #1251, we ran a systematic sync/async parity review — comparing each module of the upcoming async port against its sync counterpart line by line, with the sync implementation as the source of truth. Along the way the review surfaced a handful of small pre-existing issues in the sync code. This PR fixes the six of them we could verify, each with a focused regression test. They are all small and independent; happy to split or drop any of them if you'd prefer.

1. Blue/Green: two message keys referenced by logger calls are missing from the bundle. SuspendConnectRouting.WaitConnectUntilCorrespondingHostFound is logged with a format argument (blue_green_plugin.py:504); Logger.debug's formatted path has no missing-key guard, so with DEBUG logging enabled a switchover in the suspend-until-corresponding-host path raises NotInResourceBundleError. BlueGreenStatusProvider.AllGreenHostsChangedName (no-arg, :1791) was silently skipped. Both keys are added.
2. Limitless: _is_login_exception discarded its result (limitless_plugin.py:524-525 had no return), so the if self._is_login_exception(e): raise guards in the connect/retry paths never fired and login failures were retried like transient errors. It now returns the classification verdict. Note: this is a small behavior change — auth failures now surface immediately instead of burning the router retry budget, which appears to be the callers' intent.
3. Limitless: the synchronous get-routers path queried a stale connection. After context.set_connection(...) replaces a dead/None monitoring connection, the router query still used the pre-reconnect local variable (:505-510). It now queries the fresh connection.
4. Session state: the transfer-handler reset was a silent no-op. reset_transfer_session_state_on_switch_func assigned a nonexistent attribute (states/session_state_service.py:99-101 — reset_transfer_session_state_on_switch_callable instead of transfer_session_state_on_switch_callable), so a registered custom transfer handler could never be cleared.
5. Plugin service: raw message key raised. plugin_service.py:475 raised the literal string "PluginServiceImpl.CouldNotDetermineCurrentHost" instead of the resolved message (the key exists in the bundle).
6. Aurora initial connection strategy: the retry deadline never used its timeout property. Both _get_verified_writer_connection and _get_verified_reader_connection computed the deadline from OPEN_CONNECTION_RETRY_INTERVAL_MS (default 1 s), and OPEN_CONNECTION_RETRY_TIMEOUT_MS (default 30 s) was never read anywhere — effectively capping the whole retry budget at one interval. The deadline now uses the timeout property.

The five limitless retry tests were updated to stub is_login_exception as False (they exercise generic retryable errors — previously the discarded verdict masked the distinction).

Verification: mypy (237 files), flake8, isort, and pytest ./tests/unit -Werror (1,123 passed) all clean.

Proposed CHANGELOG entry (under ### :bug: Fixed):

Six small fixes found during the sync/async parity review: missing Blue/Green log message keys (crash under DEBUG logging during switchover), Limitless login-exception classification discarded and stale monitoring-connection read, session-state transfer-handler reset not clearing the handler, a raw message key raised by the plugin service, and the Aurora initial connection strategy never reading open_connection_retry_timeout_ms. (PR #1256)

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.