Make iterators hold a reference to the cache

[Malkiz223]

Closes #69

The iterator types kept a raw cursor and a generation counter, but no reference to the cache object, so a cache that nothing else held was freed while the walk was still running, and the walk read freed memory. Now every iterator type holds a Py<PyAny> to the cache: keys, values and items take slf instead of &self and store it. __traverse__ is there so the cycle "cache holds the iterator as a value, iterator holds the cache" stays collectable; __clear__ is not needed, the cache's own __clear__ breaks that cycle.

No answers change, and speed does not change either: creating an iterator stays at 37 ns.

Two tests in the shared mixin, so they run for all seven types: walking a cache that nothing else holds, and a cache that holds its own iterator still being collected. The walk runs in a child process on purpose: reading freed memory takes the whole run down instead of failing a test. Without the patch it fails for Cache and LRUCache and passes for the other five, which is what reading freed memory looks like.

1011 passed, 6 skipped, 5 runs out of 5. Those runs have #68 applied on top: on plain main the suite freezes in almost every run because of #67, which this branch does not touch.

[awolverp]

You pointed out a good point. Thanks for the solution. It just seems like there’s a problem
on PyPI