[CUDA][CUBLAS] Fix BF16 reduced precision reduction note in Numerical accuracy docs (#101884)
Fixes #100966
Ref #101044
Align implementation and documentation. (This is what's previously missed from the above issue and PR)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/101884
Approved by: https://github.com/eqy, https://github.com/ezyang
