Use GPU-friendly rounding

[pxl-th]

• Use GPU-friendly rounding.
  Otherwise we create boxed value that is used for reporting exceptions which triggers hostcalls on AMDGPU.
• Use Int for indexing (as it is default anyway) instead of converting to UInt32.

Fixes #573.

PR Checklist

• Tests are added
• Documentation, if applicable

[mcabbott]

Is there any way this could have a test?

And are you reasonably sure that changing UInt32 to Int won't hurt CUDA cases? I don't know why that was written that way, it looks odd to me, but haven't investigated original PR etc.

[pxl-th]

Is there any way this could have a test?

Test that would check we are not launching hostcalls on AMDGPU?

And are you reasonably sure that changing UInt32 to Int won't hurt CUDA cases? I don't know why that was written that way, it looks odd to me, but haven't investigated original PR etc.

I think that was me being obsessed with keeping Int32 indexing back then. It might save us a few registers, but beyond that there should be no visible performance impact.
We are still producing Int64 with that code, because @index returns them, so does size(..) and so on.