Improve CPU float64 exp SIMD accuracy

[chrismicah]

Summary

This is narrow partial progress for #3047: it improves CPU float64 accuracy for exp by adding a double-precision SIMD polynomial path in mlx/backend/cpu/simd/math.h.

It intentionally does not claim to fix the other fp64 unary ops mentioned in #3047 (sin, cos, erf, erfinv). I checked the prior closed PR #3058 and kept this patch aligned with the maintainer feedback there: this is not a scalar element-wise std::exp fallback.

Evidence

Pre-change, CPU float64 exp([0.1, 0.5, 0.9, 1.25]) showed fp32-scale error from another local checkout at 0.32.0.dev20260621+5abdd04b:

• max absolute error: 2.783627603974992e-07

Post-change, after rebuilding the local editable MLX package from this branch:

• python -m pip install -e . --no-build-isolation passed
• CPU float64 exp max absolute error: 4.440892098500626e-16
• softmax-like fp64 exp-normalization max absolute error: 1.3877787807814457e-17
• python -m pytest python/tests/test_double.py -q passed: 10 passed
• python -m pytest python/tests/test_ops.py -q -k 'exp or softmax' passed: 7 passed, 134 deselected

I also added a C++ ops_tests.cpp regression that checks float64 exp against std::exp at 1e-12 tolerance. I could not run the C++ doctest binary locally because the editable build tree did not include the C++ test executable / doctest headers, so this is draft until that is covered by CI or a local C++ test build.

Notes

• Related issue: [BUG] [CPU] Some ops fallback to fp32 implementation even fro fp64 #3047
• Prior attempt checked: Fix double precision for exp, sin, cos, erf, erfinv on CPU #3058
• This is a draft until C++ test coverage is confirmed.

[zcbenz]

We have been experimenting with a google/highway based SIMD implementation (#3019) and the SIMD code will go though a thorough rewrite, so while this is nice to have, we probably won't accepts PRs on it recently.