Changes

Refactor kfold_model_validation to use SQLite as primary data store

Previously, all metrics were stored in pandas DataFrames and written to CSV after each patient, causing O(n²) slowdown for large cohorts (e.g. patient 2400 took significantly longer than patient 1).

Key changes:

• Introduced SQLite database (results.db) as the primary data store, with WAL mode enabled for crash-safe writes
• Replaced per-patient CSV writes with SQLite INSERT/UPDATE operations
• CSV files are now exported once after all folds are complete, rather than after each patient
• Added resume support: existing database is reused on restart, with missing CSV files regenerated from the database if needed
• Replaced init_from_file (which re-read the full CSV on every patient) with init_from_db, which queries only the relevant rows from SQLite, eliminating the O(n²) bottleneck
• Extra metric columns are added via ALTER TABLE when new metrics are introduced after the initial run
• compute_fold_average now reads all extra metric columns present in the CSV, not just those in the current config, preserving metrics from previous runs

Fix early-exit logic in compute_patient_extra_metrics to correctly skip already-computed metrics

The original early-exit check had two bugs:

• Used x == x to detect computed values, which evaluates to True for None (since None == None is True in Python), causing the function to return early with None values instead of recomputing them.
• Called get_optimal_class_extra_metrics multiple times redundantly.

The new implementation:

• Checks explicitly that values are not None using v is not None and v == v
• Only checks values relevant to the configured metric spaces (patientwise/objectwise) rather than assuming both are always present
• Stores the result in a local variable to avoid repeated function calls