changefeedccl: descriptor leases leaked in rowFetcherCache and unbounded wait in changeAggregator.close

**Summary:**

Two changefeed paths hold resources across a failure that skips their release. Part of #174944.

**Findings:**

- `cdcevent.rowFetcherCache.refreshUDT`: the cache holds a node-lifetime `*descs.Collection` as a struct field and calls `ReleaseAll` only on the success path; both error returns leave the leases held. Because the collection is never torn down, the leaked refcounts persist for the process lifetime, which stalls DDL on the affected descriptors cluster-wide rather than expiring with a lease. Reachable from every changefeed with UDT columns, and it fires on ordinary errors today. It also reuses a per-transaction object across transactions.
- `changefeedccl.changeAggregator.close`: a bare `<-ca.kvFeedDoneCh` receive with no timeout, as the second of eight unguarded cleanup steps. If the kvfeed goroutine is wedged, flow cleanup hangs; for a remote flow that parks the stopper waiter task, which blocks node drain.

**Next Steps:**

- [ ] `defer collection.ReleaseAll(ctx)` in `refreshUDT`
- [ ] Bound the `kvFeedDoneCh` receive with a timeout and a loud log
- [ ] Consider whether the remaining seven cleanup steps in `changeAggregator.close` should each be independently guarded

Epic: none

Jira issue: CRDB-68127

Start at cdcevent.rowFetcherCache.refreshUDT and changefeedccl.changeAggregator.close, tracing both error paths and the kvFeedDoneCh cleanup. Done means descriptor leases are released after failures, close cannot wait forever on a wedged kvfeed, and the timeout emits a loud log; review whether the other seven cleanup steps need independent guards.

Written by the indexing model from the issue text.