Remove usage of torch.autograd.profiler_legacy for benchmarks

[anmyachev]

Part of #1905

Closes #2150

Since #1905 is blocked for now it's probably good idea to change profiler to xpu elapsed_time in order to continue investigate performance degradation caused by removing IPEX import. I think we've come to terms with the increase in absolute performance numbers at this point (see #1905 (comment)).

[anmyachev]

https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/10746078971/job/29806332779 is failed at the moment

UPD: Fixed, see the last run: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/10760495839/job/29838516905

[whitneywhtsang]

Any performance impact?

[anmyachev]

Any performance impact?

Let's take results (TFLOPS) from https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/10760495839 and https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/10763825651.

Note: unfortunately the script from #2060 doesn't work for this PR

Summary (old version vs new version):

1. Softmax perf:
   • Absolute numbers:
     • On average difference: -17% (Triton), -14% (XeTLA)
     • Maximum difference: -44% (Triton), -48% (XeTLA)
   • Ratio (Triton/XeTLA) numbers:
     • On average difference: 2%
     • Maximum difference: 10%
2. Attn perf:
   • Absolute numbers:
     • On average difference: the same (<1 %) (Triton), -93% (XeTLA)!!!
     • Maximum difference: the same (Triton), -98% (XeTLA)!!!
3. Matmul perf:
   • Absolute numbers:
     • On average difference: the same (< 1%) (Triton)
     • Maximum difference: 10% (Triton)

[anmyachev]

We can try another method. It looks like the previous measurement method only included the sycl kernel time, and the new approach (elapsed_time) also includes everything that was done to prepare this kernel, for example, a large number of allocations in the case of "Attn" benchmark. If we remove them from the measurement, the average difference becomes only 16%!!! instead of 93% (for "Attn").

I believe that by working in this direction it is possible to achieve an acceptable deterioration in absolute numbers of performance while maintaining the ratio (within some acceptable limits). This could be an acceptable solution until bugs in the other method are fixed. This also unlocks the ability for us to benchmark on platforms that only work with Upstream PyTorch.

New CI run on this PR with last changes (with moving allocations from the measured function): https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/10783177036/job/29904547469.

@etiotto @whitneywhtsang @alexbaden thoughts?