feat(gui): add Language dropdown to Advanced settings panel

[pretyflaco]

Summary

Whisper's auto-detect on sparse / quiet audio (long opening silences, short initial utterances) is unreliable and can mis-classify English as ja/zh/ko, producing pages of CJK hallucinations interspersed with correctly-transcribed English fragments.

A real Blink dev-sync recorded on 2026-05-05 was rendered useless this way: the meeting was English but the .txt / .srt / .json / .summary.md / .pdf artifacts all came out in Japanese hallucinations like:

[00:01:46] Openoms: Linuxと同じです。私はセーブとクライアントについて...
[00:04:06] Jonas:   Rust SDKは、Rust SDKは、Rust SDKは、Rust SDKは、Rust SDKは...

Letting the user set the meeting language up-front via the GUI avoids the failure mode without forcing them to drop to the CLI.

Changes

• New fourth row in the GUI Advanced expander: a Language combobox.
• Values: auto (default) plus en, de, fr, es, tr, fa, it, pt, nl, ja, zh, ko, ar, ru (alignment-models registry plus common Whisper languages).
• Selection writes back into self._transcribe_kwargs["language"] immediately so the next recording's transcription picks it up.
• The CLI --language plumbing (in place since PR Add support for Apple Silicon and PyTorch device configuration #4) is unchanged; this PR only surfaces it in the GUI.

Tests

• New test_launch_passes_explicit_language in tests/test_gui.py.
• Existing test_launch_defaults_pass_none_through extended to assert language=='auto' is the default.

Repair workflow for affected sessions

meet transcribe <session_dir> --language en --no-summarize
# then re-apply speaker labels via meet label or apply_labels

The 2026-05-05 session this PR was motivated by has been repaired manually with --language en --mixdown mono followed by apply_labels(...).

Followups (separate PRs)

• Stash recovery: chore(sync): rebase pull and push-only-when-ahead.
• Stash recovery: fix(label): handle sensitive condenser mics in YOU detection.
• Optional: log Whisper language-detection confidence so low-confidence detections are visible (feat: expose ASR backend, torch device, and MLX model in GUI (#5) #11 / B.2 from the investigation).