feat: dual vocab embeddings

[schennam714]

---

Summary

• Separate embedding tables for melody and chord tokens - Models now use dedicated melody_embed and chord_embed layers instead of a single unified embedding, allowing each modality to develop task-appropriate representations
• Removed runtime logits masking - Output layers now project directly to chord_vocab_size, making chord-only output an architectural constraint rather than a runtime mask
• Fixed eval decoding bug - Chord predictions now correctly decode using chord_id_to_token instead of unified_id_to_token
• Preprocessing now saves separate vocab files - vocab_melody.json and vocab_chord.json are saved alongside vocab_unified.json

Changes

Models (models/offline.py, models/online.py):

• Offline: Custom encoder_embed for melody input, T5's internal embedding for decoder (chord) tokens
• Online: melody_embed + chord_embed selected via is_melody boolean mask
• Both models output logits over chord_vocab_size only (no masking needed)
• Removed _ChordOnlyLogitsProcessor and logits_mask buffer

Data (data/base.py, data/online.py):

• BaseDataset now loads separate vocab_melody.json and vocab_chord.json
• Added _unified_to_chord_id() and _unified_to_melody_id() conversion methods
• OnlineDataset uses base class vocab sizes (removed redundant recalculation)

Eval (eval/offline.py, eval/online.py):

• Fixed vocab decoding: chords use chord_id_to_token, melody uses melody_id_to_token

Preprocessing (scripts/preprocess.py):

• Added melody_vocab.save() and chord_vocab.save() calls

Tests:

• Added semantic validation tests ensuring chords decode as chord tokens
  (not melody)
• Updated all model tests for new dual-vocab API

Test plan

• All 54 unit/integration tests pass
• Ruff linting passes
• mypy type checking passes
• Train offline model and verify convergence
• Train online model and verify convergence
• Run eval on trained checkpoints

---