Use TensorList in group_index_select_dim0 (#1884)

Summary:
Pull Request resolved: #1884

Replace variadic template in `group_index_select_dim0` by packing all
tensors into `TensorList` and passing the `TensorList` to autograd
function. Return output needs to be tensors of the same size as
input specified in the forward functions and can be returned as
type of variable_list. This improves performance and lifts
limitations of using variadic template that needs to be instantiated
for each group size and limits max group size of 55.

Reviewed By: sryap

Differential Revision: D47488358

fbshipit-source-id: aa2c23e1b038743d54e5878ff9e8eb6ccd27e852