Implement caching properly

[skirpichev]

Currently caching not works in PyPy and rely on implementation details of the CPython.

I believe, there could be a more correct way is to resurrect the deallocated object. One possibility is using the tp_finalize slot. Though, it seems to be more slow, see #300.

[skirpichev]

I think this can be closed.

Freelist's seems to be a standard optimization in the CPython nowadays, used also in the Cython (since cython/cython@232d3ca). It does work on the GraalPy, but doesn't have much sense here. It seems, same is true for PyPy: see pure-Python benchmark in 3bcfcb4.