PERF: short-circuit array_equivalent and equals comparisons

[jbrockmendel]

Summary

• Add short-circuiting Cython functions for array equality checks in pandas/_libs/lib.pyx:
  • array_equivalent_float: single-pass NaN-aware float comparison, replacing the 4-temporary-array expression ((left == right) | (isnan(left) & isnan(right))).all()
  • array_equivalent_bytes: memcmp-based comparison for int/bool/datetime arrays, replacing np.array_equal
  • has_nans/all_nans: short-circuiting replacements for np.isnan(arr).any()/.all()
• Wire these up in array_equivalent, _array_equivalent_float, _array_equivalent_datetimelike, and the equals methods on DatetimeLikeIndex, MultiIndex, MaskedArray, and Categorical
• For non-contiguous inputs, fall back to the original numpy expressions with no regression

array_equivalent with dtype_equal=True, 10^6 elements:

dtype	equal	early mismatch
float64	2.0x	800x
int64	1.1x	380x
datetime64	1.1x	235x
bool	1.8x	60x

DataFrame.equals on 1000x1000:

dtype	equal	early mismatch
float64	298 us	4 us
int64	165 us	4 us

closes #32339

Test plan

• All pre-existing equals tests pass (frame, series, index, dtypes, arrays, internals)
• Tested correctness for all dtypes: float32/64, complex64/128, int, bool, datetime, timedelta
• Tested edge cases: empty arrays, NaN-heavy arrays, different shapes, non-contiguous, F-contiguous, strided slices
• Verified no regression for adversarial inputs (non-contiguous falls back to numpy, F-contiguous 2D uses transpose trick)

🤖 Generated with Claude Code

[rhshadrach]

lgtm, just a question

[rhshadrach]

Is it worth adding asserts for the common unsafe ones (float/complex/object I think)?

[jbrockmendel]

i dont think so; the caller is responsible for gating

[rhshadrach]

Thanks @jbrockmendel