CUTLASS is integrated into TVM

[hwu36]

apache/tvm@541f9f2

In this first commit, Turing and Ampere FP16 tensor core GEMMs are added into TVM.

Great work and thank you very much, TVM community!!!

[hwu36]

Now cutlass batched GEMM is integrated, too.

[masahi]

Convolution kernels are also integrated. Results and analysis on end to end models are available at apache/tvm#9746.

It looks like TVM + cutlass is faster than TVM + cudnn. I'm not claiming that "cutlass is faster than cudnn on my gpu", it's just that the way we use cudnn is apparently not performing as well as it should be.