Keep the DRb connection pool alive after socket errors

[OskarEichler]

Contains stale-connection liveness and eviction-close errors inside the pool worker, closing broken entries when possible so later callers do not deadlock behind a dead worker.

Verified with the full 122-test / 547-assertion suite and focused stale/eviction error models.

[kou]

Could you also share your real world problems instead of just opening auto generated PRs to many repositories?

[OskarEichler]

This was generated from a broad dependency audit, not from a specific Songstats production incident. I can explain the hypothetical dead-worker path, but I cannot honestly provide a real-world failure that motivated it. That context is not strong enough for this change, so I am closing the PR.