Fix floor division with infinite operands

[tandede]

Summary

• add a dedicated single-output FloorDivide primitive for floating-point inputs on CPU, Metal, and CUDA
• match Python and NumPy for infinite operands, including finite values divided by oppositely signed infinities and infinite numerators
• keep the quotient returned by divmod consistent with floor_divide for the same edge cases
• preserve the integer path and explicitly reject unsupported complex inputs
• support compilation, vmap, and autodiff for the new primitive, with regression coverage for eager, compiled, and divmod execution

Root cause

Floating-point floor division was implemented as floor(a / b). That is not equivalent to Python floor division at infinity: floor(1 / -inf) preserves negative zero instead of returning -1, while floor(inf / 3) remains infinite instead of becoming NaN.

The fix uses a dedicated primitive so the exceptional cases can be corrected in one backend pass without computing an unused remainder or materializing an intermediate quotient array. The CPU implementation retains a SIMD fast path when neither input contains infinities.

Closes #4317.

Validation

• pre-commit run --all-files: all hooks passed
• CPU C++ suite: 251 test cases and 3,358 assertions passed
• python/tests/test_ops.py: 162 passed, 1 skipped
• all 49 combinations drawn from NaN, +/-Inf, positive and negative finite values, and signed zero matched NumPy values and sign bits
• full CPU-only Python discovery: 801 passed and 40 skipped; the remaining 18 MPS DLPack cases require Metal and failed because this local build intentionally uses MLX_BUILD_METAL=OFF

On one million finite float32 elements, the dedicated CPU primitive reduced the local median evaluation time from 0.248 ms to 0.202 ms across seven 20-iteration samples (about 1.23x). Metal and CUDA compilation and runtime coverage are left to CI because those toolchains are not available locally.

Checklist

• 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 (no API documentation change is required)