Re-organize the backward split code generation (pytorch#1871)

Summary:
Pull Request resolved: pytorch#1871

- Move the determination of `kMaxVecsPerThread` and `kThreadGroupSize` in `embedding_backward_split_template.cu` into its own macro

- Update `embedding_backward_split_template.cu` to hardcode values for kMaxVecsPerThread and kThreadGroupSize for the experimental optimizers case

Reviewed By: sryap

Differential Revision: D47354614

fbshipit-source-id: 85e947734d929d86339a348e3fb4bda1eeb24dce