Support non-standard bools in CUDA masked_scatter (#79391)
The failure comes from within the `exclusive_sum` used to calculate
the selection index. So, I merged the code paths for `uint8_t` and
`bool` masks together, then use `cub::TransformIterator` to get a
proper `bool` value for the result.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/79391
Approved by: https://github.com/mruberry, https://github.com/mikaylagawarecki
