Clean up internals, fix bugs, add tests + CI

[bleugreen]

Summary

Remove internal/unused code, fix several bugs, deduplicate utils, and add a basic test suite with CI.

Changes

Removed unused code

• Deleted model/train.py, dataset/, bench/, requirements.txt
• Deleted model/infer.py after extracting load_cnn_model() into predictor.py

Bug fixes

• hcqm.py: Changed device='cuda' default to device=None, resolved with get_device() — no longer crashes on CPU-only machines
• load_cnn_model(): Added weights_only=True to torch.load()
• batch_infer.py: Guarded torch.cuda.empty_cache() with is_available() check
• batch_infer.py: Fixed indentation, removed stray pass

Code deduplication (utils.py)

• split_audio() is now the core implementation
• load_and_split_audio() is a thin wrapper that loads then delegates
• split_audio raises ValueError on too-short audio instead of silently returning None
• load_and_split_audio catches exceptions with a clear error message

Package structure

• Added __init__.py to model/ and audio_proc/
• Fixed circular import between predictor.py and batch_infer.py
• Organized imports across touched files

Packaging (pyproject.toml)

• Added missing requests runtime dependency
• Added optional dev dependencies (pytest, ruff)
• Added ruff + pytest config

Tests + CI

• 7 tests covering BPM conversion, audio splitting, predictor instantiation, and end-to-end prediction
• GitHub Actions workflow: lint with ruff, run pytest on PRs and pushes to main

Verification

• pip install -e ".[dev]" — installs cleanly
• ruff check src/ — all checks passed
• pytest — 7/7 pass
• python -m deeprhythm.infer --help — CLI works