PERF: bypass factorize in axis=1 EA reduction fastpath (GH-56903)

[jbrockmendel]

Summary

• The EA-backed axis=1 reduction fastpath in DataFrame._reduce flattens the frame into a 1D EA and groups by row indices to simulate a transpose+reduce. The row indices passed to Series.groupby were already in factorized form (np.tile(np.arange(nrows), ncols)), but the standard groupby pipeline re-factorized them.
• Call arr._groupby_op directly with the pre-known ngroups and ids, skipping the redundant factorize and the surrounding groupby pipeline. idxmin/idxmax get an inline conversion from flat positions to column positions plus the same skipna=False NA-detection that GroupBy._idxmax_idxmin was raising before.
• Closes PERF: pd.BooleanDtype in row operations is still very slow #56903.

Benchmark (Apple M1 Pro)

import pandas as pd, numpy as np

shape = 250_000, 100
pd_mask = pd.DataFrame(np.random.randint(0, 2, size=shape)).astype(pd.BooleanDtype())
%timeit pd_mask.any(axis=1)
%timeit pd_mask.all(axis=1)
%timeit pd_mask.sum(axis=1)

op	before	after
any	256 ms	117 ms
all	277 ms	112 ms
sum	204 ms	74 ms

Small-frame numbers (10k x 10, BooleanDtype) drop from ~1.1 ms to ~0.35 ms across any/all/sum/min/max/mean.

cProfile on the 250k x 100 case showed factorize at 43% of wall time pre-change; that line item is gone post-change.

Test plan

• pandas/tests/reductions
• pandas/tests/frame/test_reductions.py
• pandas/tests/extension
• pandas/tests/arrays/masked
• pandas/tests/groupby/aggregate
• pre-commit run on changed files
• mypy / pyright on pandas/core/frame.py

🤖 Generated with Claude Code