Implement disk-based caching for stem separation

Implement Spleeter support and RAM-aware stem backend selection

Implement Apple Silicon optimized demucs-mlx backend for stem separation

stem_render: accept max_iter param, use it for refinement iterations"

Make render_full_set_with_stems accept progress flexibly"

Fix stem_render function signature and add stem-based analysis"

Patch render_full_set to use demucs stem separation + stem_mixer\n\nAdds import at top and overrides render_full_set to:\n1. Run demucs htdemucs on each track\n2. Use stem_mixer.mix_stems for surgical drum/bass transitions\n3. Falls back to filter-based mixer if demucs fails"