feat: regression default metrics — RMSE first, MAPE only for positive targets

[rxavier]

Summary

Implements P0-3 from the ML defaults improvements plan (docs/ml-defaults-improvements.md).

Changes

• Default regression metrics become ["neg_root_mean_squared_error", "neg_mean_absolute_error", "r2"].
  • RMSE is now the primary metric (first column), used as the ranking/ensemble sort key.
  • neg_mean_squared_error was outlier-dominated; neg_median_absolute_error dropped.
• neg_mean_absolute_percentage_error is only included when the target is strictly positive. MAPE epsilon-floors the denominator in sklearn, silently producing astronomical values when y contains zeros.
• get_target_info() now reports whether a regression target is positive (positive key).
• Bumped scikit-learn floor to >=1.4.0 (required for neg_root_mean_squared_error).

Tests

• Defaults: RMSE first, no MAPE on a mixed-sign target.
• Strictly positive target includes MAPE.
• Zero-containing target scores cleanly (no MAPE column, no NaNs).
• Updated the two shape tests that hardcoded the old 4-metric count.

All 244 tests pass; ruff + format clean.