Benchmark time spent in GIL and GC

[gjoseph92]

It would be valuable to measure what percentage of worker time is spent blocked on the GIL, and in GC. There is plenty of discussion about the performance impact of both of these beloved Python features, so having hard numbers to measure them would be valuable:

• Worker event loop performance hampered by GIL/convoy effect dask/distributed#6325
• Are reference cycles a performance problem? dask/distributed#4987

I think it would be possible to get the GIL percentage from py-spy out of the box. GC support is in progress benfred/py-spy#389.

Not something we'd take on immediately, just an idea for a useful thing to measure in the future for things like dask/distributed#6889.