fix(metal): preserve subnormal float values when casting to bool (#4205)

[reckylurker]

Fixes #4205

Proposed changes

In Metal Shading Language (MSL), casting a floating-point number directly to bool via static_cast<bool>(x) evaluates in float registers, which flushes subnormal numbers to zero in hardware. So static_cast<bool>(subnormal) ends up evaluating as 0.0f != 0.0f -> false.

This was affecting .astype(mx.bool_) (in copy.h) as well as any(), all(), and other reductions whenever U = bool.

To fix this, introduced a mlx_cast<U, T> helper in utils.h that checks the raw bit pattern for floats (ignoring the sign bit, e.g. (as_type<uint32_t>(val) & 0x7FFFFFFF) != 0). This checks the exponent and fraction bits as raw integers so subnormals don't get flushed by Metal hardware units, while still keeping -0.0 as False.

updated copy.h, reduce_all.h, reduce_col.h, and reduce_row.h to use mlx_cast<U>.

Added test_subnormal_bool_cast in python/tests/test_ops.py to check astype(bool), any(), and all() on float32 and bfloat16 subnormals. All Python and C++ tests pass!

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 (if needed)

[zcbenz]

In cuda backend there is a similar utility named cast_to, I think we can use the same name here.

[zcbenz]

float16 does not seem to need this, should add a test.

[reckylurker]

yes, looks like half values are not affected in MSL. Thanks for the suggestion.