Fix expm1 for complex inputs on the CPU backend

[obchain]

Proposed changes

Fixes #3759.

The CPU expm1 used std::expm1, which has no complex overload, so for a complex input it operated on the real part and discarded the imaginary part (e.g. expm1(0.5 + 0.7j) returned expm1(0.5) + 0j).

Added a complex branch to expm1 in base_simd.h that uses the defining identity exp(z) - 1, matching the complex handling already present for log1p and log2. The real and low-precision (Accelerate / fp16) paths are unchanged.

Before:

>>> mx.expm1(mx.array([0.5 + 0.7j]))
array([0.648721+0j], dtype=complex64)   # == expm1(0.5), wrong

After it equals exp(z) - 1.

Verified the complex forward now matches np.exp(z) - 1 and that the (already correct) Expm1::jvp agrees with finite differences for complex inputs; real and fp16 results are unchanged. Added a complex case to test_expm1.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)