Add dataset metadata JSON

Add train/test splits (test=1500 stratified by severity)

Rename eval_heldout_metadata.json to pathological_speech .json

Add evaluation metadata JSON

Add held-out evaluation set (unseen speakers, severity metadata)

initial commit