Add patient-level leakage checks for dataset splits

[tippered1-debug]

Summary

This PR adds helper utilities to audit patient-level leakage across dataset splits.

Healthcare datasets often contain multiple samples per patient. If the same patient appears in both train and validation/test splits, evaluation metrics may be inflated because the model is partially evaluated on patients already seen during training.

The new helpers inspect the actual samples in each split and verify that patient IDs are disjoint.

Changes

• Add get_patient_ids(...) to extract patient IDs from datasets, subsets, sample collections, or ID collections.
• Add check_patient_disjoint(...) to return a report with split counts and patient overlaps.
• Add assert_patient_disjoint(...) to raise a clear error when patient overlap is detected.
• Export the helpers from pyhealth.datasets.
• Add focused tests for patient-level splits, sample-level leakage detection, conformal splits, and missing patient_id errors.

Tests

• python -m unittest tests.core.test_patient_disjoint -v