Make it such that external_splits specification can point to a patient_splits.parquet file or a prior splits.json file from MEDS-extract to match the cohort.

[mmcdermott]

Right now, if you point external_splits to a prior dataset's splits.json file, it will treat the shard name as part of the split. This should be fixed such that you can point to a single "splits" file and have it reload the right splits, not the shards part.

Tagging @prenc for tracking

My current thoughts as to what should change about this:

• splits.json should be renamed to .shards.json (note it is being made a hidden file). It should rarely be used. This also conforms with Pipelines should automatically determine shards from the input directory rather than relying on the splits.json file. #129 about how this should be standardized.
• external_splits should be made to work with dataframe files (like patient_splits.parquet)
• external_splits should throw a warning if it seems to be given a shards.json file and default to collapsing splits down to standardized names (though this should be controllable with an option in the stage_cfg).