Create Benchmark from MIMIC IV Ext Direct

Involves table merge, data exploration and transformation. Requires CITI specimen license.

Remarks:
- to get the reports, stay_ids encoded in json file names must be extracted and used to join on original MIMIC IV (v3.1) data files
- due to original MIMIC IV tables for which join is executed use either BigQuery on Google Cloud or pre-filter tables by stay id before joining locally
- two tables are needed: `hosp/noteevents` and `hosp/admissions`
- hard part: spans must be matched to text positions in report
- ICD-10 codes must be inferred from diagnostic text e.g. HFpEF, could use foundational model for pre-processing
- will modify factory to support spans instead of sentence id later
- for now only report chunking as most fine-granular resolution is supported

Implementation:

- inherit from benchmark interface, ie, provides implementation of all exposed functions like `def get_train_true` returning a DataFrame` with columns `['report_id', ’span', 'y_true’]`
- see `codiesp.py` as an example, though way more complex due to original source was in Spanish and

References and Resources:

- https://physionet.org/content/mimic-iv-ext-direct/1.0.0 (visible after license agreement)