Accept both ptb/ptb_xl and mimic/mimic_iv as splitter keys

[TonyChen06]

TLDR: I think the HuggingFace datasets (tested for the relevant datasets) don't actually split patient-wise. Please double check yourself but for me it doesn't work. If so, this is probably the root cause. vv

Splitter._PATIENT_EXTRACTORS is keyed on "ptb" and "mimic", which only matches inputs whose ecg_path segments are data/ptb/... or data/mimic/.... Inputs that use the longer form (data/ptb_xl/..., data/mimic_iv/...) hit _PATIENT_EXTRACTORS.get(...) → None, so _patient_id returns None for every row, every row gets appended to loose as a singleton group in split_dataset, and the train/test split silently degenerates to row-level random shuffling. The post-split assertion doesn't catch this because the comprehensions building train_pids / test_pids filter out None with if splitter._patient_id(...), leaving both sets empty so the intersection is trivially empty. Empirically, willxxy's HF parquets — which use the long form — show 99.9% / 91% / 40% / 57% test-ECG-in-train overlap on ecg-qa-ptbxl-250-2500, ecg-qa-mimic-iv-ecg-250-2500, ecg-instruct-pulse-250-2500, ecg-instruct-45k-250-2500 respectively (leakage rate scales with questions-per-ECG). Fix is to add aliases so both path conventions resolve to the same extractor.