[FR] Easy attaching perf from within the library

**Is your feature request related to a problem? Please describe.**

Sometimes benchmarks have a heavy setup which affects counters/recordings.
It would be nice to easily skip heavy initialization when using perf.

**Describe the solution you'd like**

folly benchmarks has `--bm_perf_args` which will start perf after doing a first iteration of a benchmark.
Running the first iteration of the benchmark allows to skip initialization of statics.

For example in folly benchmarks I can run `--bm_perf_args=record` or `--bm_perf_args=stat`.

That works ok but isn't very generic. Maybe you know something more generic?

**Describe alternatives you've considered**
It is possible to do this through external tools but is very annoying.