Harden prediction load and multiclass CV against invalid/empty input

[gbeane]

Addresses two robustness review comments from the feature/multiclass → main PR.

1. prediction_manager — invalid/corrupted prediction records

io.load() raises ValueError when a saved record has an invalid schema (e.g. class_names present but probabilities are 2-D instead of (n_id, n_frames, n_classes)). _load_prediction_record only caught KeyError/FileNotFoundError, so load_multiclass_predictions() would propagate the error instead of falling back to empty outputs.

• io.load is now wrapped so a ValueError (corrupt/invalid record) returns empty outputs, consistent with the missing-file behavior.
• The deliberate mode-mismatch guard (multi-class record requested as binary or vice-versa) still raises — that's a logic error, not a corrupt file — and now runs after the load rather than inside the same catch.
• Fixed test_load_multiclass_predictions_invalid_shape_returns_empty, which previously passed for the wrong reason: it wrote the group as __multiclass__ while the loader looks it up via to_safe_name(...) = multiclass, so it hit a KeyError and never exercised the invalid-shape path. It now writes the correct safe group name and genuinely exercises the ValueError fallback.

2. cross_validation — empty labels_by_behavior

In multi-class mode _prepare_cv_labels called merge_labels(...) unconditionally; merge_labels raises on an empty dict (no labeled frames yet), so CV crashed instead of being skipped like the binary path.

• _prepare_cv_labels now short-circuits on empty labels_by_behavior, returning an empty label array so the caller finds no valid splits and skips CV gracefully.

Tests

• New test_multiclass_cv_skips_when_no_labels (CV skipped, no raise).
• Fixed/strengthened the invalid-shape prediction test.
• tests/classifier + tests/project (298) pass; ruff clean.