[GradualGating] support better k value change (#41557)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41557
- add new learning rate functor "slope"
- use "slope" learning rate in gated_sparse_feature module
Test Plan:
buck test dper3/dper3/modules/tests:core_modules_test -- test_gated_sparse_features_shape_num_warmup_tensor_k
buck test caffe2/caffe2/python/operator_test:learning_rate_op_test -- test_slope_learning_rate_op
Reviewed By: huayuli00
Differential Revision: D22544628
fbshipit-source-id: f2fcae564e79e1d8bcd3a2305d0c11ca7c0d3b3c
