Fix patient merging on null patient_id tables (per-partition reset_index)

[fbonc]

Problem
For tables configured with patient_id: null (one patient per row), the code assigned patient_id via df.reset_index(drop=True) then df.index. In Dask reset_index resets per partition, so rows at the same position in different partitions collide onto the same id and get silently merged into one "patient". This corrupts several shipped configs (covid19_cxr, medical_transcriptions, clinvar, eICU hospital) and defeats split_by_patient / split_by_patient_conformal, whose entire purpose is to prevent cross-patient leakage.

Fix
Assign a globally-unique id with a cumsum (global across partitions) instead of the per-partition index, in both base_dataset.py and fhir/base.py.

Tests
Added a regression test for each path that builds a null-patient_id table across multiple Dask partitions and asserts every patient_id is unique. Both fail on the old code and pass on the fix.