Update Attention op to support attention mask for GPT-2 (#4330)
* Support another two format of mask_index input: 2D attention mask, or 1D mask index with end and start positions.
* Update dynamic axes of gpt2 with past state
* Update script to fuse model with attention mask