Accumulate CPU half-precision sums in float32

[sofinvalery]

Proposed changes

Accumulate CPU float16 and bfloat16 sum reductions in float32, while preserving the output dtype. This fixes precision loss in ops that use sum().

Fixes #3326.

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)

[sofinvalery]

Made it inline. Feels cleaner.

[zcbenz]

Allocating a new array would have heavy performance penalty, the correct way would be refactoring strided_reduce/contiguous_reduce to accumulate in float32 rather than the output type.