Add batch optimization, listening-note preference learning, and waveform preview

[MLyte]

Motivation

• Implement the requested feature set: batch processing, A/B listening notes / preference learning, and a waveform preview in the GUI.
• Persist lightweight listening history so user notes can influence future ranking decisions.
• Give a quick visual overview of the source with a small waveform image to help review candidates faster.

Description

• Added CLI commands optimaster optimize-batch ... to run optimization over multiple files and optimaster add-note ... to save listening ratings (changes in src/optimaster/cli.py).
• Introduced PreferenceStore to persist notes and compute per-preset bias, saved to renders/preferences.json, and a PreferenceStore API (src/optimaster/preferences.py).
• Integrated preference bias into scoring inside EngineService.optimize so saved listening notes adjust candidate scores and append a scoring reason when bias is applied (src/optimaster/service.py).
• Added FFmpeg-based waveform preview generation render_waveform_preview and wired it into the GUI to display a waveform image after file selection, plus rating controls and a "Save listening note" action in the recommended candidate panel (src/optimaster/ffmpeg.py, src/optimaster/gui.py).
• Updated docs/backlog to mark these features implemented (README.md, TASKS_MASTER.md) and added unit test tests/test_preferences.py covering preference persistence and bias direction.

Testing

• Ran PYTHONPATH=src pytest -q and all tests passed: 5 passed.
• Verified compiled syntax via python -m py_compile for the modified modules and it succeeded for src/optimaster/gui.py, cli.py, service.py, ffmpeg.py, preferences.py.
• Note: a plain pytest -q run initially failed in this environment because the package root was not on PYTHONPATH, so tests were re-run with PYTHONPATH=src and succeeded.

---

Codex Task