Inspect Triton CPU launch latency

[xuzhao9]

walltime ms: 0.013 ms

triton-lang/triton#3619

We don't observe there is a regression, but it would be good to track the overhead and compare it with tilelang/tvm-ffi.
AMD metrics are pretty noisy, so we are not showing them here.

[xuzhao9]

Looks like there is no regression, so we are closing this for now. Need to add tilelang launch latency.

[xuzhao9]

Reopen as we observe Triton launch latency is ~50% slower than tilelang: #669

We also want to add tvm-ffi with cutedsl, as well as Kineto trace for the CPU launch latency.

[xuzhao9]

Profiled the triton kernel launch process at #674