using float type to do the computation of norm reduce for cpu half and bfloat16 dtype (#95166)
As the title, we should use a higher dtype to compute norm reduce for half and bfloat1 dtype.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95166
Approved by: https://github.com/peterbell10, https://github.com/jgong5, https://github.com/ngimel, https://github.com/lezcano
