Fix ListQueues: iterator leak, unbounded CQL round-trips, and PageState after Close

[mykaul]

What changed?

1. Set explicit PageSize for Cassandra queue ReadMessages queries — ensures CQL paging
   is respected even when a LIMIT clause is present.

2. Add upper bound to GetTaskQueuesByBuildId result set — caps results at 10,000 to
   prevent unbounded iteration over CQL pages.

3. Fix iterator leak in ListQueues — three error paths inside the scan loop
   (getQueueFromMetadata, GetPartitionForQueueV2, getMessageCountAndLastID) returned
   without calling iter.Close(), leaking the gocql iterator (server-side cursor and
   connection resources).

4. Add CQL round-trip cap to ListQueues — with ALLOW FILTERING, Cassandra may return
   under-filled pages because it scans a fixed number of partitions per page and then
   post-filters. The old single-page fetch could return fewer results than requested even
   when more exist. The new code loops over CQL pages (up to maxListQueuesPages = 10)
   until the requested PageSize is filled or all pages are exhausted, returning a valid
   NextPageToken if the cap is hit.

5. Capture PageState() before iter.Close() — the old code called iter.PageState()
   after iter.Close(), which is incorrect per gocql semantics.

6. Pre-allocate queues slice with make([]QueueInfo, 0, PageSize) to avoid repeated
   slice growth.

7. Export TemplateGetQueueNamesQuery for test coverage consistency with other template
   constants.

Why?

ListQueues uses ALLOW FILTERING because the partition key is (queue_type, queue_name)
but we filter only on queue_type. Cassandra handles this by scanning a fixed number of
partitions per CQL page and then discarding non-matching rows. This means a single page
fetch can return 0 matching rows even when many exist — the caller must fetch additional
pages to fill the requested page size.

The iterator leak is a resource correctness issue: on any error during row processing, the
gocql iterator was not closed, leaking a server-side cursor and potentially a connection.

How did you test it?

• built
• run locally and tested manually — against both Cassandra and ScyllaDB
• covered by existing tests (all TestCassandraQueueV2Persistence tests pass)
• added new unit test: ListQueuesGetQueueNamesQuery — verifies that when the
  ALLOW FILTERING query itself fails, iter.Close() error is properly surfaced as
  *serviceerror.Unavailable with the QueueV2ListQueues operation name
• added failingQuery.PageState mock method to support the new multi-page loop

Potential risks

• The maxListQueuesPages = 10 cap means that in extreme cases (very large number of
  non-matching partitions), ListQueues may return fewer results than PageSize with a
  valid NextPageToken. This is a deliberate trade-off to prevent unbounded CQL queries.
  Callers that need all results should paginate using NextPageToken.

[CLAassistant]

All committers have signed the CLA.