Fix tensor core instruction shape

[nlaanait]

Major Changes

1. Change MMA_K Instruction shape for the "tensor_core" benchmark to conform with shape+stride of Inputs
   ref: https://github.com/NVIDIA/cutlass/blob/df18f5e4f5de76bed8be1de8e4c245f2f5ec3020/include/cute/atom/mma_traits_sm80.hpp#L147-L160

Minor Changes

1. Registered a "tensor_core" benchmark to benchmarks

[BradLarson]

Is this left over from debugging?

[nlaanait]

correct- was planning on constraining the inputs dtype to float32 (to make the tensor_core code path happy) but realized it's probably more robust to come up with a mapping of dtype to supported MMA instructions to generalize the examples to other input types.

[nlaanait]

removed.
LMK what you think of the mapping above, happy to contribute it in a separate PR.

[BradLarson]

Thanks, let me check and see what we want to do to generalize this over multiple datatypes.

[BradLarson]

Thanks for the addition and fix! Had one comment on a possibly extraneous precondition, but otherwise looks good. I can update the example in the MAX repo with the same, to match.