feat: OHE min_frequency knob to collapse rare categories

[rxavier]

Summary

Implements P1-5 from the ML defaults improvements plan (docs/ml-defaults-improvements.md).

Changes

• Added ohe_min_frequency parameter to PoniardPreprocessor (default 5).
  • int: absolute category count; float: fraction of samples; None: keep every category.
• Default OHE is now OneHotEncoder(drop="if_binary", handle_unknown="ignore", sparse_output=False, min_frequency=5).
• Rare categories collapse into sklearn's infrequent_sklearn bucket, composing correctly with handle_unknown="ignore" (unseen categories still map to all-zeros).
• Disabled the knob in the two existing tests that assert exact OHE output on 5-row toy data (where every category is rarer than 5), so they keep testing their intended behaviors.
• Documented the new default in the preprocessing guide.

Tests

• Rare categories collapse into the infrequent column; None preserves every category; float thresholds work as fractions.

All 251 tests pass; coverage 89.4% (gate 85%); ruff + format clean.