[torch_ops] Benchmarks

[amd-eochoalo]

We would like to have some microbenchmarks defined at the same abstraction level as torch operations. The infrastructure behind these microbenchmarks should follow a similar design (and reuse code wherever possible) as the torch operation correctness tests. Other considerations have been collected somewhat loosely here:

• iree-run-module/iree-benchmark-module may have limited applicability here since timing it will also include time spent on CPU and not on kernel dispatches.
• iree-benchmark-executable would be ideal to use, as it is target agnostic, but it currently lacks GPU timestamps.
• torch.profiler brings torch as a dependency and we would like these tests to not depend on torch.
• rocprofv3 target specific.
• tracy would need a debug build of iree-run-module.

At the moment, the work being done focuses on using rocprofv3 and will generalize to use iree-benchmark-module/iree-benchmark-executable for other targets.