Bump the generation when a read promotes an LRUCache entry

[Malkiz223]

Closes #75

LRUPolicy::get moved the entry to the back of the list without bumping the generation, so a live iterator kept walking the relinked list and the process died. Now a promoting read bumps the generation, and a live iterator raises the same RuntimeError it already raises after an insert. A hit on the most recent entry moves nothing, so it does not bump either: iterating is still safe around it, and the pointless unlink-relink it used to do is gone.

PolicyExt::get takes shared now, the same way entry already does; only the LRU policy uses it, the others ignore it.

No answers change, eviction order stays exactly the same. A promoting hit pays one atomic increment, 3 ns on my machine; a hit on the most recent entry and misses cost what they cost before.

The test runs the walk in a child process on purpose: without the fix it dies instead of failing.