cuda: extended MMF_ROWS_PER_BLOCK

[zhang-hui-yulo]

extended MMF_ROWS_PER_BLOCK in mmf to more than warp_size, just keep MMF_ROWS_PER_BLOCK to the old value as I don't do the performance tuning.

Has been tested MMF_ROWS_PER_BLOCK = 64 on my 3080, no enough shared memory when MMF_ROWS_PER_BLOCK = 128

[am17an]

You can use this PR as a base for other PRs, as such there is no use for this in it's current form right, i.e. this path is not exercised?

[JohannesGaessler]

Please keep this in the PR for WMMA support, I cannot evaluate these changes in a vacuum.

[zhang-hui-yulo]

You can use this PR as a base for other PRs, as such there is no use for this in it's current form right, i.e. this path is not exercised?

Yep, the path is not exercised unless you set MMF_ROWS_PER_BLOCK to 64, as the performance might be increased or decreased in test-backend-ops due to the different shape, so I just keep the old 32 value.