[compile_trace] Add compile time Kineto trace

[xuzhao9]

Generates Kineto trace focusing on compile time so that we can peek into compile time analysis.

Fixes #117

Test plan:

$ python run.py --op softmax --num-inputs 1 --input-id 0 --metrics compile_trace --only triton_softmax

Internal link only: https://www.internalfb.com/intern/perfdoctor/trace_view?filepath=tree%2Ftritonbench%2Fcompile_time.json&bucket=tc_bench_ci

With autotuning:

[facebook-github-bot]

@xuzhao9 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[xuzhao9]

cc @fywkevin @Jokeren This is a Python side view of the Triton compilation process, note that Kineto has no observation on the C++ code.

[xuzhao9]

For autotuning, the Triton's autotuner has some GPU idleness gaps:

[Jokeren]

This is a Python side view of the Triton compilation process, note that Kineto has no observation on the C++ code.

I think it's fine to have such a coarse-grained view from profilers.
MLIR can get you a very detailed view of time spent on each compiler passe, but I think it may only provide insights for compiler engineers.