Use tp_finalize slot to manage cache

[skirpichev]

Benchmark	ref	patch
collatz0(97)	87.5 us	103 us: 1.17x slower
collatz0(871)	131 us	154 us: 1.17x slower
collatz0((1<<128)+31)	706 us	836 us: 1.18x slower
collatz1(97)	93.2 us	105 us: 1.13x slower
collatz1(871)	141 us	159 us: 1.13x slower
collatz1((1<<128)+31)	757 us	857 us: 1.13x slower
Geometric mean	(ref)	1.15x slower

With collatz.py benchmark from #263 on CPython 3.14 with --enable-optimizations.

Bigger picture, including CACHE_SIZE=0 cases:

Benchmark	ref-nocache	ref	patch-nocache	patch
collatz0(97)	129 us	87.4 us: 1.47x faster	140 us: 1.09x slower	103 us: 1.25x faster
collatz0(871)	193 us	132 us: 1.46x faster	210 us: 1.09x slower	154 us: 1.25x faster
collatz0((1<<128)+31)	1.01 ms	707 us: 1.43x faster	1.11 ms: 1.10x slower	832 us: 1.22x faster
collatz1(97)	129 us	93.6 us: 1.38x faster	143 us: 1.11x slower	106 us: 1.22x faster
collatz1(871)	195 us	140 us: 1.39x faster	215 us: 1.10x slower	159 us: 1.22x faster
collatz1((1<<128)+31)	1.02 ms	758 us: 1.35x faster	1.15 ms: 1.12x slower	857 us: 1.19x faster
Geometric mean	(ref)	1.41x faster	1.10x slower	1.23x faster