Often at most one instance is allocated at any given time, so this improves performance quite significantly in common cases. Uses the freelist feature introduced in #19316.

This speeds up self check by about 0.5% (measured 100 samples). Speeds up this microbenchmark that gets close to the maximal benefit by about 90%:

def f(x: int) -> int:
    def inc(y: int) -> int:
        return y + 1
    return inc(x)

def bench(n: int) -> None:
    for x in range(n):
        f(x)

from time import time
bench(1000)
t0 = time()
bench(50 * 1000 * 1000)
print(time() - t0)