Support masks in make_block_ptr for GEMM

[minjang]

The recent GEMM code using make_block_ptr delivers very good performance, which is similar or better than PyTorch for many sizes. But it currently can't handle mask. @ienkovich do you have a plan to support it?