Fit describe-code metadata on training split only

[florian6973]

What changed

• fit describe-code frequencies from data/train/**/*.parquet for split-sharded MEDS datasets
• fall back to filtering unsharded data with metadata/subject_splits.parquet
• fail clearly when train-only fitting cannot be guaranteed
• isolate and reduce only train-only cache files, preventing stale tuning or held-out cache entries from affecting fitted metadata
• make Polars row orientation explicit when converting frequency dictionaries
• add regression coverage for threshold leakage, held-out-only vocabulary entries, subject-split filtering, and missing split metadata

Why

meds-tab-describe recursively scanned all Parquet files under input_dir. On split-sharded datasets, tuning and held-out covariates therefore influenced metadata/codes.parquet and downstream feature filtering such as min_code_inclusion_count.

This is label-free held-out preprocessing leakage: evaluation covariates could affect the fitted feature vocabulary even though labels were not used.

Impact

Preprocessing metadata is now fitted only on training subjects and then reused for all splits. A code with 8 train occurrences and 5 held-out occurrences now has fitted count 8 and will not pass a minimum count of 10.

Validation

• uv run pytest -q: 82 passed
• targeted regression tests: 3 passed
• Ruff check and format check passed
• git diff --check passed

Fixes #166