Defer heavy ML imports to fix slow tray startup

[Copilot]

Launching the tray (python recorder.py tray) was taking ~5 minutes to display the icon because torch, pyannote.audio, and faster-whisper were imported at module load even though none are needed to draw the tray UI.

Root cause

recorder.py had module-top from diarizer import Diarizer, and diarizer.py had module-top import torch + from pyannote.audio import Pipeline. Importing pyannote.audio transitively pulls in torch, torchaudio, pytorch-lightning, speechbrain, asteroid-filterbanks — minutes of cold-start cost on Windows.

Changes

• diarizer.py — replaced top-level torch / pyannote.audio imports with a lazy _ensure_pyannote_imports() helper. Only invoked when the pyannote backend is actually selected (i.e. HF_TOKEN is set) or when PyannoteDiarizer._load_pipeline runs. Backend-selection semantics in _init_backend are preserved.
• recorder.py — moved from diarizer import Diarizer / from transcriber import Transcriber into Recorder.__init__, and reordered main() to dispatch the tray subcommand before constructing Recorder. The tray app already builds the Recorder lazily inside its recording thread on first "Start Recording" click, so user-facing behavior is unchanged — the ML stack just loads on demand instead of at process start.

# recorder.py main()
if args.command == "tray":
    from tray_app import run_tray
    run_tray(config)
    return

recorder = Recorder(config)  # ML imports happen here, not at module load

After the change, import recorder completes in ~17 ms with no torch / pyannote / faster-whisper modules loaded.