[PyTorch] use div instead of mul when calculating sampling probability (#70001)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/70001
multiplying inversion of `kLowProb` instead of division which uses less expensive `mul` instead of `idv`
Test Plan:
Before
{F682076291}
After
{F682076323}
Reviewed By: robieta
Differential Revision: D32608440
fbshipit-source-id: 7851317a0f7e33813f2bd7a152e5e7f4b5c361b4