Clearing a cached artifact and queueing its file aren’t atomic #467

#428 made retention clear the record and queue the file in pending_deletes in one transaction. Three other places still do it in two steps:

• DiscardArtifact (internal/database/queries.go) clears the record, commits, then calls QueuePendingDelete.
• UpsertArtifact commits the new storage path, then queues the old one separately.
• The LRU loop in evictBatch (internal/server/eviction.go) clears the record, deletes the file, and only queues it if the delete fails.

A crash or a failed second write between the steps leaves a file in storage that no record points to and nothing will ever delete. I'd move each pair into one transaction, the same way ClearExpiredArtifact does now.