* [Eager] Support final_state_adam when argument grad (position 1) is selected_rows

* Remove needless code

* Add adam_dense_param_sparse_grad kernel