* move workspace memory-allocation to PyTorch

* refine the code based on the comments

* remove unnecessary options

* remove bsz from set_seq_len function