gh-130701: create new PGO task based on benchmarks

[nascheme]

As described in the issue, using the unit tests as the PGO task has some problems. This PR creates a new PGO task by taking benchmarks from the "pyperformance" suite and adjusting them (slightly) to work better as a PGO task. I had a couple objectives when moving the benchmark code over:

• use only benchmarks that don't have external dependencies
• prefer benchmarks that exercise the most heavily used parts of Python
• adjust code so that execution time is mostly spent doing "real work" rather than on timing related things
• adjust the loops or iterations so that each task takes around 0.1 to 1 seconds, roughly

There are some potential issues in adding this new PGO task:

• adding the additional code to the cpython repo is not ideal. It is approximately 500 kB of data added to the Lib/test/pgo_task folder. We could put the code in a separate repo or PyPI package but I think we prefer that Python can be compiled without additional external dependencies.
• the code duplication between pyperformance and Lib/test/pgo_task is not ideal. OTOH, the pyperformance benchmarks don't see a lot of code changes (in order for benchmark results to be stable) and so keeping the code up-to-date should not be too much of a problem.
• the current task set doesn't cover as much code as the unit test based PGO task did. I think this can be resolved over time by adding more modules to the pgo_task folder to cover those code paths.
• It's possible that using the new PGO task will result in worse optimization of the Python executable, for real work-loads. However, I think if that happens we should adjust both the pyperformance and the pgo_tasks so those execution patterns are better represented. We are focusing a lot of optimization efforts based on what pyperformance results say and so those benchmarks should also be a good representation of real-world programs. If not, we should improve them.

Benchmark results from pyperformance.

• Issue: Using the unit tests as the PGO task has problems #130701

[brandtbucher]

I agree with the root issue, but I'm skeptical of using our benchmarks as a profiling task (it feels too much like "gaming" them).

As a data point, I tried using pyperformance as a PGO task a couple of years ago, and it made things about 4% faster (according to pyperformance, of course): faster-cpython/ideas#99 (comment)