fix: improve metrics reliability and edge-case handling

[AhmedAli58]

Summary

This PR addresses three meaningful quality issues in the metrics layer, with regression tests.

Problem statements and fixes

1. regression_metrics_fn was order-dependent because kl_divergence mutated intermediate arrays used by later metrics.

• Fixed by using local safe copies for KL computation so mse/mae remain unaffected by metric ordering.

2. multilabel_metrics_fn mutated y_pred while computing ddi, which could break subsequent metrics in the same call; it also returned only ddi_score instead of the requested ddi key.

• Fixed by using a local pred_labels list (no mutation), returning ddi, and preserving ddi_score for backward compatibility.

3. ranking_metrics_fn lacked robust validation and failed unclearly on edge cases (invalid k_values, empty score output).

• Added input validation for empty args and non-positive k_values, narrowed import error handling, and added an explicit error when evaluator output is empty.

Impact

• Eliminates order-dependent metric behavior.
• Improves reliability for mixed metric requests in multilabel tasks.
• Produces clearer, earlier failures for ranking edge cases.

Tests

• Added tests/core/test_metrics_quality.py with regression coverage for all three issues.
• Verified with:
  • python -m py_compile pyhealth/metrics/regression.py pyhealth/metrics/multilabel.py pyhealth/metrics/ranking.py tests/core/test_metrics_quality.py
  • HOME=/tmp python -m unittest tests.core.test_metrics_quality -v