Keep borrowed packs open after cache eviction

[bojanz]

Several code paths close a Pack at the moment it leaves the object-store cache: _add_cached_pack closes the pack it replaces, _enforce_packed_git_limit closes the least-recently-used pack, _remove_pack closes on removal, and _update_pack_cache closes packs whose files disappeared.

Closing at eviction time is unsafe when the store is shared: packs and get_raw hand out references to cached Pack objects, so another consumer — a concurrent request in a threaded/greenlet server, or an iterator like iter_unpacked that is still mid-pack — may be reading from the pack when eviction closes it under its feet. The reader then fails on a closed file even though the pack file itself is still perfectly valid on disk.

This changes eviction to release ownership instead of closing: the pack is dropped from the cache, and the underlying files are closed by Pack.__del__ once the last borrower is done with it. Since that hand-off is intentional, __del__ no longer emits the "unclosed Pack" ResourceWarning for packs released this way — explicit leaks still warn as before.

The new test exercises the LRU-limit path: a pack evicted by packed_git_limit while an iter_unpacked iterator is mid-stream stays readable until the iterator completes.

[bojanz]

The Windows failures were real: test_packed_git_limit_evicts_lru_packs started leaving an evicted pack's .idx open into teardown, so rmtree failed with WinError 32. Pushed a fix.

Unsure how useful test_no_reference_cycle truly is, considering that we already had a test that failed.

[bojanz]

The last commit could be a followup PR as well if you prefer it that way.

One behavioral note: work and errors that previously happened at the call, most notably the full sort in sorted_entries() and the checksum/length validation on first self.data access in iterobjects(), now happen at the first next().

[bojanz]

Drafted cause I want to take another look at this once we land #2334