Fix half->float case of softmax backward when inner_size is not 1 (#30838)
Summary:
Fixes https://github.com/pytorch/pytorch/issues/30572
That unit test is tested to fail with master and success with this PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30838
Differential Revision: D18841066
Pulled By: ngimel
fbshipit-source-id: 86a7ccdb3016c98d62dd0946daff101704cd1f68