Implement WMMA

[pxl-th]

Very primitive matmul WMMA kernel benchmarking:

Size               WMMA-Ptr    rocBLAS      Naive       WMMA    rocBLAS
                       (ms)       (ms)       (ms)     TFLOPS     TFLOPS
-----------------------------------------------------------------------
256x256x256           0.043      0.056      0.178       0.78       0.60
512x512x512           0.061      0.106      0.857       4.40       2.52
1024x1024x1024        0.670      0.739      1.110       3.21       2.91
2048x2048x2048        1.384      1.128      4.206      12.41      15.23
4096x4096x4096        8.606      5.571     15.224      15.97      24.67

Closes #560.
Closes #483.