New microbenchmark on the functools.lru_cache(lambda: 42) function using my 3 optimizations on _PyType_GetModuleByDef():

* commit d4aaa34798f0dd8402f412e2aa9d6fa2d6cff5fa
* commit 760da626ff4124e1344fd8b7dbeb83b2c4b7c12c
* commit cdad2724e6f7426372901cc5dedd8a462ba046a6
* I didn't pick Raymond's optimization for this benchmark, so I can still use _functools for my benchmark.

$ python3 -m pyperf compare_to py39.json master.json 
Mean +- std dev: [py39] 38.8 ns +- 0.5 ns -> [master] 39.2 ns +- 0.9 ns: 1.01x slower

The _PyType_GetModuleByDef() overhead in _functools is now about +0.4 ns, it's better than my previous measurement before optimization: +5.7 ns (37.5 ns +- 1.0 ns -> 43.2 ns +- 0.7 ns).

These timings are really tiny, it's really hard to get reliable timing even with CPU isolation. For example, two measurements on Python 3.9:

* Old measurement: 37.5 ns +- 1.0 ns
* New measurement: 38.8 ns +- 0.5 ns

I guess that for timings under 100 ns, the PGO build is no longer reliable enough. Moreover, the std dev is around to 1 ns on 40 ns.