Robust OpenML dataset selection: retries, resume preference, and binary validation

[charlesmartin14]

Motivation

• Improve robustness of the random OpenML binary dataset selection so the notebook can resume prior choices, avoid repeatedly failing dataset loads, and ensure selected datasets are actually binary.

Description

• Change default REUSE_LAST_MODEL behavior to True and use resume_dataset_uid to optionally try the prior dataset id first when not forcing a fresh start.
• Add MAX_DATASET_SELECTION_ATTEMPTS and build a shuffled candidate list limited to that many dataset ids to avoid long or infinite selection loops.
• Introduce a selection loop that attempts to load_dataset for each candidate, catches loader exceptions, validates that y is binary, records failures, and selects the first valid candidate or raises a clear RuntimeError if none succeed.
• Move configure_checkpoint_paths so paths are set after a valid selected_dataset_uid is found and initialize X, y, meta defensively before the selection loop.

Testing

• No automated tests were added or executed for this change.

---

Codex Task