Speed up MathFeatures reductions with NumPy

[Sudaiskhan1133]

Summary

• use vectorized NumPy reducers for common MathFeatures row-wise aggregations
• preserve pandas semantics for string and NumPy callable reductions, including ddof and missing values
• retain the existing pandas agg fallback for custom functions, uncommon reducers, and nullable extension dtypes
• document and test the optimized behavior while keeping the benchmark comparable

Why

MathFeatures currently routes common reductions through row-wise DataFrame.agg, which adds substantial per-row Python overhead. On a local 100,000-row x 5-feature benchmark using sum, mean, std, min, and max, the legacy logic took 30.49 seconds while the optimized path took 0.0159 seconds (about 1,920x faster).

Compatibility

• matches pandas results for missing and all-missing rows
• string std and var retain pandas' ddof=1 behavior
• NumPy callable semantics remain version-aware
• custom callables and nullable dtypes continue through the legacy pandas fallback

Validation

• 36 MathFeatures tests passed on pandas 2.3.3 and pandas 3.0.5
• 184 creation tests passed on both pandas versions
• 2,062 tests passed in the initial full suite; the 51 dataset-cache failures were rerun with a writable cache as part of 163 passing affected-file tests
• flake8 passed
• mypy passed for the changed module
• changed module reached 99% statement coverage with no missed statements
• Sphinx documentation build passed with warnings treated as errors

Closes #576