[BUG]: `math.copysign` doesn't restrict it's dtype to float about mojo

[JoeLoser]

Agreed. As seen in the LLVM docs, only FP types are accepted. I just added a use of constrained in this function so you get a slightly nicer error message — at least explaining that only FP types are permitted.

[soraros]

@JoeLoser come think about it, copysigndoes seem to have well defined meaning for integer types, and are quite convenient in generic code. Maybe we should have implementation for them as well, just not lowered into a single LLVM intrinsic.

[JoeLoser]

@JoeLoser come think about it, copysigndoes seem to have well defined meaning for integer types, and are quite convenient in generic code. Maybe we should have implementation for them as well, just not lowered into a single LLVM intrinsic.

Ha, funny that you mention that. I just reworked the PR a few minutes ago to define it for integral types too with the same rationale. :)

[soraros]

@JoeLoser Lol! For signed types only or both? What about booleans? How was zero treated? Maybe having to answer these questions is why they are not generally available (not in NumPy or PyTorch even, I just checked). I feel very conflicted right now.

[gryznar]

btw, why not copy_sign to align naming?