Fix NaN-aware top/bottom N parity across engines

[emiliolugo]

Summary

• make top_n/bottom_n NaN-aware in records.nb without dtype promotion, preserving deterministic tie order for non-NaN data
• align Rust top_n_mapped_mask/bottom_n_mapped_mask semantics with Numba for NaN-containing columns (including n > n_valid and all-NaN cases)
• add regression/parity coverage for NaN selection, overflow behavior, and tie-order determinism

Backward compatibility

• non-NaN behavior is preserved; changes apply to NaN-containing inputs where selection is now explicit and engine-consistent

Test plan

• uv run --python 3.11 --extra test pytest tests/test_records.py -k "top_bottom_n_mask_with_nan or top_bottom_n_mask_with_all_nan_column or top_bottom_n_mask_tie_order"
• uv run --python 3.11 --extra test pytest tests/test_engine.py -k "top_n_mapped_mask_nan or bottom_n_mapped_mask_nan or top_bottom_n_mapped_mask_nan_overflow" (Rust-gated tests skip when rust package is unavailable)
• uv run --python 3.11 --extra test pytest

Co-authored by Cursor

[polakowo]

Merged, thanks!