fix deadlock between in-progress builds & delete-crate/release

[syphar]

I'm quite sure this is the reason why we sometimes stop consuming from the index. And it's when we get deletes while we have an in-progress build.

What happens today (simplified):

• Builder txn: locks queue row (FOR UPDATE SKIP LOCKED) and keeps it locked while build runs.
• Delete txn: locks releases/metadata rows, then tries DELETE FROM queue ... and waits on builder’s queue lock.
• Builder’s build path (different DB txn/conn) may try touching releases and waits on delete txn.
• Now each side waits on the other -> deadlock.

For now the simple solution is to make the delete wait for the build to be finished. I added a configurable timeout.

I want to postpone the better solution to when we have something event-based between crates.io & docs.rs, where I then could easily NACK some events, and handle them later.

[GuillaumeGomez]

Nice, thanks!