Add SongEval and shared model caching

[ftshijt]

Summary

• add production-ready SongEval support with pinned upstream assets, optional dependencies, local/offline configuration, audio validation, and stable result keys
• make --cache_folder consistently configure utterance and corpus metrics in both scorer entry points
• share model caches by backend for Hugging Face, Whisper, ESPnet, Torch Hub, and NeMo while preserving explicit metric cache_dir overrides
• route cache settings through existing model-backed metrics including ARECHO, NeMo WER, NOMAD, PAM, Qwen2-Audio, Qwen2.5-Omni, speaking rate, SQUIM, and VAD
• document the shared-cache layout and add focused regression coverage

Motivation

Versa exposed cache options in several places, but the main scorer did not apply --cache_folder to ordinary utterance metrics and some pretrained-model loaders ignored metric cache configuration. This could duplicate large checkpoints or place downloads in user-specific default locations. SongEval also needed a reliable setup path for both its predictor and MuQ encoder.

This change makes a mounted cache root reusable across runs and machines while keeping unrelated metric artifacts isolated.

Validation

• 132 passed, 1 skipped across metric definitions, base metric/pipeline coverage, SongEval, shared-cache routing, NOMAD, PAM cache forwarding, and speaking-rate tests
• Black checks pass for all changed Python files
• bash -n tools/install_songeval.sh
• git diff --check
• wheel build with pip wheel . --no-deps --no-build-isolation
• real SongEval + MuQ CPU inference:
  VERSA_RUN_REAL_MODEL_TESTS=1 python -m pytest test/test_metrics/test_songeval.py::test_songeval_real_model_inference -vv -s
  (1 passed in 43.04 seconds)

The real-model test remains opt-in because it downloads approximately 8.5 GB
of MuQ assets in addition to the SongEval checkpoint.