[pytorch][glow][NNPI] Using int32 as indices for embedding_bag operators (#45878)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/45878
Support int32 as indices and offsets for embedding_bag_byte|4bit_rowwise_offsets, to avoid costly casting operators such as `aten::to`.
Currently we don't make the assumption that indices and offsets should be the same type, which should not be a problem since downstream fbgemm supports either cases.
Test Plan:
```
buck test mode/dev caffe2/test:quantization -- --stress-runs 100 test_embedding_bag
```
Reviewed By: radkris-git
Differential Revision: D23854367
fbshipit-source-id: 6758a4252b36a7fe2890f37d38d66f20651e850e
