Lock the current store file and close failed acquisition handles

[OskarEichler]

Summary

Acquire a lock on the current store inode, retrying if an atomic save replaced it while acquisition was pending. Close abandoned/failed handles through ensure. This also closes write-mode descriptors on flock failure.

Reproduction and verification

A deterministic writer opened the old inode, paused before flock, then resumed after another writer atomically replaced the file. Two increments previously produced 1; this patch produces 2. A bounded five-writer/100-update run ended at 23 with atomic writers and 28 with mixed writers on the original, versus 100 for fast/atomic/mixed modes after this patch. IOError/Interrupt lock fixtures also preserve exceptions and close their handles.

• External focused corpus: 8 cleanup checks plus deterministic and concurrent writer cases verified on this isolated branch under rbenv Ruby 4.0.6.
• Existing upstream RBENV_VERSION=4.0.6 rbenv exec rake test: 13 tests / 26 assertions, zero failures before and after.
• All files and injected failures are confined to scratch storage; no production operations.
• No repository tests added or modified under the contributor's task constraint. Other Ruby versions and native Windows were not executed locally.

Breaking-change notes and limitations

The inode check adds filesystem work to every successful acquisition, including fast mode, because another cooperating writer may use atomic mode. Randomized paired local benchmark (11 measured samples after two warmups, 500 transactions each): cleanup-only consumer 12.057ms fast / 69.504ms atomic; inode-retry variant 13.116ms / 72.145ms. The consumer audit has therefore withheld the inode-retry portion under its performance policy; descriptor cleanup alone is adopted. No speed improvement is claimed. Native Windows, NFS and mixed old/new implementations are not verified; same-path cooperating patched writers were tested.