fix(soniox): surface STT server errors

[he-yufeng]

Fixes #5854

Summary

• turn Soniox STT error_code / error_message frames into APIStatusError
• preserve the existing final transcript / usage flush before surfacing the error
• let APIError propagate through the receive loop and _run instead of being swallowed by broad exception handlers
• add a unit test for an error frame without tokens

Tests

• python -m pytest tests\test_plugin_soniox_stt.py -q -p no:cacheprovider
• python -m ruff check livekit-plugins\livekit-plugins-soniox\livekit\plugins\soniox\stt.py tests\test_plugin_soniox_stt.py
• python -m ruff format --check livekit-plugins\livekit-plugins-soniox\livekit\plugins\soniox\stt.py tests\test_plugin_soniox_stt.py
• python -m py_compile livekit-plugins\livekit-plugins-soniox\livekit\plugins\soniox\stt.py tests\test_plugin_soniox_stt.py
• git diff --check

[he-yufeng]

I checked the failing test-stt job. The Soniox cases from this PR are skipped in that run because SONIOX_API_KEY is empty; the actual failures are in other hosted providers (openai, fireworksai, aws) plus the fork-comment 403 while posting the generated test summary.

The PR-specific local checks from the branch are still the focused Soniox regression tests plus repo lint/type checks. I don't see evidence in the CI log that the Soniox error-frame change caused the current hosted-provider failures.

[tinalenguyen]

thanks you for the PR!