Add reproducible chronological validation support

[sergioald]

Summary

This PR adds reproducible chronological train/validation/test splitting support for time-dependent datasets.

The implementation is designed to make temporal validation explicit, deterministic, replayable, and resistant to accidental leakage across split boundaries.

What is included

• deterministic chronological train/validation/test splits
• fraction-based and explicit split boundaries
• numeric and datetime coordinates
• NumPy datetime64 support
• timezone-aware pandas timestamps normalized to UTC
• support for point samples and [start_time, end_time] intervals
• strict exclusion of intervals crossing split boundaries
• optional sample-count gaps / embargoes between partitions
• immutable NumPy index arrays
• deterministic validation manifests
• SHA-256 fingerprints of ordered time coordinates
• manifest replay with dataset-order/content verification
• validation of malformed, overlapping, or inconsistent boundaries
• documentation and regression coverage

Reproducibility

ValidationSplitManifest records the split definition together with a fingerprint of the ordered temporal coordinates.

apply_split_manifest() verifies the fingerprint before reproducing the split, so reordered or changed datasets are rejected rather than silently producing a different validation partition.

Scope

This PR intentionally focuses on a small reusable chronological-validation primitive.

It does not add:

• rolling-origin or walk-forward validation
• expanding-window validation
• grouped/station-level splitting
• benchmark execution
• dataset-specific validation logic

Those can be considered separately if useful.

Validation

The implementation was checked locally before submission.

• focused validation suite: 112 passed
• complete repository test suite: 674 passed
• git diff --check: passed
• Ruff/static checks: passed

No unrelated changes are included.