Support collecting cProfile data from a benchmark

[mdboom]

While this wouldn't be useful for absolute timing (benchmarking), it's sometimes handy to generate a profile from a benchmark to understand where the time is being spent. It's currently possible to do this by hacking up a benchmark script, but it would be better to make it as simple as:

python run_benchmark.py --profile output.profile

[mdboom]

I can't self-assign, but to be clear, I'm also offering to do this work ;)