Enable half for CUDA dense EmbeddingBag backward. (#19293)
Summary:
I audited the relevant kernel and saw it accumulates a good deal into float
so it should be fine.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19293
Differential Revision: D14942274
Pulled By: zou3519
fbshipit-source-id: 36996ba0fbb29fbfb12b27bfe9c0ad1eb012ba3c