fix(evals): restore speech when switching from text to audio sessions

[skulitom]

The eval transport now applies skip_tts on every client connection before on_client_connected, so an audio scenario can produce speech after a text-mode scenario on the same bot process. Previously, only truthy flags were sent to the LLM service; the True setting persisted when the next client omitted the flag or sent skip_tts=false.

This addresses the separate eval-harness problem reported in #5425 (comment). The original Flux/muting issue is outside this change.

Validation:

• Two regression tests exercise text -> audio -> text connections and inspect the active output setting when each connection callback runs. Both fail on main with [True, True, True] instead of [True, False, True], and pass with the fix.
• pytest tests/test_evals_transport.py tests/test_websocket_server_transport.py tests/test_evals_session.py: 120 passed.
• Repository-wide ruff check and ruff format --check pass; scoped Pyright reports no errors or warnings.

AI disclosure: OpenAI Codex prepared the code, tests, and PR text and ran the validation above. The new regression tests use mocked transport I/O; no live LLM/TTS provider evaluation was run.