Update randomness.rst (#21337)
Summary:
Following [this question on the forums](https://discuss.pytorch.org/t/reproducibility-and-performance/46504), I propose the following doc change. It clarifies that 'performance reduction' concerns the processing speed (and not the training accuracy).
Related website commit: https://github.com/pytorch/pytorch.github.io/pull/211
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21337
Differential Revision: D15622151
Pulled By: soumith
fbshipit-source-id: f0edeb20049f2ee715c400e7c57abb966864d621