[Example] warp specialization gemm for hopper

[soodoshll]

Still cleaning up the code and tuning performance.

Benchmark on H200

w/ warp specialization:

       m      n      k   name  latency (ms)      tflops
0   4096   4096   4096  torch      0.179568  765.386669
1   4096   4096   4096  tilus      0.217712  631.287910
2   4096   4096  14336  torch      0.611840  786.212620
3   4096   4096  14336  tilus      0.679744  707.672791
4   8192   8192   8192  torch      1.448576  759.029314
5   8192   8192   8192  tilus      1.771616  620.626388
6  10240  10240  10240  torch      2.862048  750.331101
7  10240  10240  10240  tilus      3.750112  572.645213

w/o warp specialization:

       m      n      k   name  latency (ms)      tflops
0   4096   4096   4096  torch      0.179392  766.137605
1   4096   4096   4096  tilus      0.221408  620.749733
2   4096   4096  14336  torch      0.612464  785.411598
3   4096   4096  14336  tilus      0.692160  694.978516
4   8192   8192   8192  torch      1.447856  759.406751
5   8192   8192   8192  tilus      1.840032  597.550281
6  10240  10240  10240  torch      2.987184  718.899036
7  10240  10240  10240  tilus      3.776592  568.630034

Benchmark on H100

Performance w/ warp specializaiton:

       m      n      k   name  latency (ms)      tflops
0   4096   4096   4096  torch      0.285840  480.824791
1   4096   4096   4096  tilus      0.342080  401.774306
2   4096   4096  14336  torch      1.050144  458.067017
3   4096   4096  14336  tilus      1.329120  361.920928
4   8192   8192   8192  torch      2.372816  463.378361
5   8192   8192   8192  tilus      3.895360  282.261878
6  10240  10240  10240  torch      5.452080  393.883370
7  10240  10240  10240  tilus      7.888288  272.236972

Performance w/o warp specialization:

       m      n      k   name  latency (ms)      tflops
0   4096   4096   4096  torch      0.286512  479.697019
1   4096   4096   4096  tilus      0.323552  424.781656
2   4096   4096  14336  torch      1.052896  456.869745
3   4096   4096  14336  tilus      1.522928  315.862823
4   8192   8192   8192  torch      2.381680  461.653800
5   8192   8192   8192  tilus      4.800304  229.050419
6  10240  10240  10240  torch      5.461200  393.225600
7  10240  10240  10240  tilus      9.372672  229.121816

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyaoding]

Thanks @soodoshll, the PR LGTM!

[yaoyaoding]

/ok to run 3507ec3

[yaoyaoding]

Consider using the sign-commits script to sign all commits, so that the ci will automatically run.

[soodoshll]

/ok to test 3507ec3

[soodoshll]

/ok to test 27a5b82

[yaoyaoding]

Thanks @soodoshll !