fix #30480 torch.normal shape checking is broken (#32243) (#33050)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33050
Following what gchanan proposed in #30480
- If the (logical) shapes of mean and std are broadcastable, we broadcast them for the output
Done in tensor iterator already.
- If the (logical) shapes of mean and std are not broadcastable and they have the same number of elements, we fall back to the old behavior (pick the shape of mean)
Done by reshape std to the same shape of mean.
- If the (logical) shapes of mean and std are not broadcastable and don't have the same number of elements, we error out.
Done by tensor iterator already.
Test Plan: Imported from OSS
Differential Revision: D19771186
Pulled By: glaringlee
fbshipit-source-id: a0b71063c7f5fdda2d4ceb84e06384414d7b4262
