[caffe2] L2 regularization for (RowWise)SparseAdagrad fusion on GPUs (#37805)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37805
Resolve the unit test failures after https://github.com/pytorch/pytorch/pull/37653
Test Plan:
```
buck test mode/dev-nosan //caffe2/caffe2/fb/net_transforms/tests:fuse_sparse_ops_test -- 'test_fuse_sparse_adagrad_with_sparse_lengths_sum_gradient \(caffe2\.caffe2\.fb\.net_transforms\.tests\.fuse_sparse_ops_test\.TestFuseSparseOps\)'
```
```
buck test mode/dev-nosan //caffe2/caffe2/fb/net_transforms/tests:fuse_sparse_ops_test -- 'test_fuse_sparse_adagrad_with_sparse_lengths_weighted_sum_gradient \(caffe2\.caffe2\.fb\.net_transforms\.tests\.fuse_sparse_ops_test\.TestFuseSparseOps\)'
```
Reviewed By: jspark1105
Differential Revision: D21395764
fbshipit-source-id: e8224a1ecbff5dce42ab732c0977de352fe98914
