fix(data): detect unusable dataset media before training

[jsugg]

Makes an incomplete or damaged dataset checkout fail early, at preparation time,
with a message that names the fix — instead of surfacing as an opaque decode
error deep inside feature extraction, hours into a run.

Git LFS pointers (_internal/utils/audio_utils.py)

A checkout without Git LFS materialized leaves ~130-byte text pointer files
exactly where the audio should be. These are perfectly readable files, so the
failure previously looked like corrupt audio rather than a setup problem.

• Detect the version https://git-lfs.github.com/spec/v1 prefix before decoding
  and raise AudioIntegrityError naming the repair (git lfs pull / checkout).
• Decode failures now raise AudioDecodeError chained (raise ... from) to the
  underlying librosa/soundfile cause, which was previously discarded.
• Both subclass OSError, so existing callers keep working unchanged. This
  typing is what lets the readiness work distinguish a sample-local decode
  failure (quarantinable) from a broken checkout (must abort).

CREMA-D validation (adapters/crema_d.py, strategies/default.py, dataset_prepare.py)

• Every candidate WAV is validated before it reaches a manifest: LFS pointers,
  empty files, files below a plausible WAV header size, and anything whose
  metadata will not decode or reports no frames.
• An empty glob is now an error — previously an empty CREMA-D tree could produce
  an empty manifest silently.
• Registered datasets are re-validated when consumed, so a checkout that
  degrades after download is still caught.
• git and git-lfs are required up front with an actionable message, rather
  than failing later once the audio turns out to be pointers.

Verification

make lint, make type (mypy + pyright clean), and the full suite — 1077 tests
— pass locally. Pre-push gate passed.