Fix elementwise unary ops giving different results for strided vs contiguous inputs (CPU) (#4163)

[AxelNoun]

Summary

• On Apple Silicon, CPU unary ops used Accelerate SIMD for contiguous inputs and scalar libm for strided views (and for the contiguous residual), so the same values could differ by ~1 ULP depending on memory layout / lane index.
• Route strided loads and the contiguous residual through the same padded SIMD kernel so results are bit-identical to the contiguous fast path.
• Commits are split: (1) strided gather, (2) contiguous residual, (3) regression tests so the residual-only fix remains submittable if bit-identity across layouts is considered out of scope (NumPy does not guarantee it either).

Fixes #4163.

Possibly related to #4161: this PR fixes the unary lane-index / length dependence shown there ( an, �rctan, etc.). Binary cases in that issue (x2, x0.5, …) share the same Accelerate-vs-libm pattern in �inary.h and are not fixed here happy to follow up the same way in �inary.h/ ernary.h if this approach is accepted. Leaving #4161 open for the binary remainder.

Validation (no local Mac GitHub Actions macos-26 / Accelerate)

	Run	Result
Before (main)	baseline	296/300 mismatches (issue repro)
After (this fix)	fix	0/300, position-independence OK, full python -m unittest discover python/tests green

Micro-benchmark (strided before vs after, float64, n=1e6, 50 iters, same runner)

The useful comparison is strided-before vs strided-after (not strided vs contiguous). Trivial ops included because gather overhead is most visible there:

Op (strided)	cols	Before (main)	After (fix)
abs	2	14.14 ms	7.91 ms
negative	2	11.86 ms	3.23 ms
tan	2	20.03 ms	9.29 ms
abs	8	11.01 ms	8.37 ms
negative	8	10.52 ms	3.38 ms
tan	8	17.28 ms	9.06 ms

No regression observed: SIMD+gather is faster than the previous pure-scalar strided path for both trivial and expensive ops on this runner.

Test plan

• Issue [BUG] mx.tan(a[:, 0]) differs from mx.tan(contiguous_copy) on float64 arrays #4163 repro script: 296/300 on main → 0/300 with fix (CI links above)
• Position-independence check from [BUG] mx.tan (and **, arctan) on float64 give different results depending on ARRAY LENGTH and LANE INDEX #4161 (unary tan)
• Full python -m unittest discover python/tests on macos-26 CPU/Accelerate
• pre-commit (clang-format / black) on touched files

[zcbenz]

Thanks for the PR but we plan to switch to use google/highway for SIMD and at the moment (#3019), so we don't accept non-critical changes on SIMD code.

[AxelNoun]

Thanks for the PR but we plan to switch to use google/highway for SIMD and at the moment (#3019), so we don't accept non-critical changes on SIMD code.

Oh okay! So much the better, tell me if there is a referenced place where the different steps that have not yet been resolved would be marked to be able to contribute to this task?