[CUDA,NVPTX] Corrected fragment size for tf32 LD B matrix.

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

Reviewed By: tra

Differential Revision: https://reviews.llvm.org/D118023