Fix cvtfp32_bf16 (#41280)
Summary:
For `Vec256<bfloat16>::blendv()` operator to work correctly, float32 -nan (0xfffffffff) must be converted to bfloat16 -nan (0xffff).
But cvtfp32_bf16 converts -nan to nan (0x7fc0)
TODO: Fix float32 +-nan conversion: i.e. float32 nan (0x7fffffff) must be converted to bfloat16 (0x7fff) nan
Closes https://github.com/pytorch/pytorch/issues/41238
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41280
Reviewed By: mruberry
Differential Revision: D23311585
Pulled By: malfet
fbshipit-source-id: 79499ce19f1ec3f6c954a874f1cd47f4ece6bdb5
