colflow: hash router output can leak disk resources when not fully exhausted

[yuzefovich]

Currently, it is possible that the disk-backed spilling queue used by hash router outputs is not closed when the hash router exits. Namely, this can occur if the router output is not fully exhausted (i.e. it can still produce more batches), but the consumer of the router output is satisfied and calls DrainMeta. In such a scenario, routerOutput.closeLocked is never called because a zero-length batch is never given to addBatch nor the output is canceled due to an error. The flow cleanup also doesn't save us because the router outputs are not added into ToClose slice. The bug has been present for several years.

Jira issue: CRDB-15161