Fall back when a topic loses every leader, count the drop, and refresh again

[koloss2001]

Summary

Port of #95 onto main.

#88 falls back only when some other partition of the topic still has a leader. If every named leader for a topic is excluded in one refresh, Candidates still returns nil and ProduceSync fails the whole batch, including healthy topics. This change treats that topic as known and hashes it onto a live agent. A topic Metadata has never returned, and a topic-level error, stay unknown so they still trigger an on-demand refresh. A partition Metadata reported with no leader (Leader below 0) does not hash and is not counted as a drop. It fails, and the existing nil-candidate path refreshes.

A refresh that excludes a leader increments warpstream_agentpool_leader_dropped_total (once per excluded leader, including the constructor) and logs the count plus the first excluded leader (first_topic, first_partition, first_node_id). The background refresh nudges another fetch. After a periodic tick that follow-up starts at once; later repeats wait out OnDemandMetadataRefreshInterval (default 1s). The constructor counts and logs, and does not nudge. Produce does not wait. AgentPool.Refresh stays ([]int32, error). The drop details are on the unexported refresh.

This also covers the broker/topic mismatch RequestCachedMetadata can return, described upstream in twmb/franz-go#1481: a leader id present in Topics but missing from that same response's Brokers now hashes onto a live agent instead of failing the batch.

Also addresses the #88 review comment on the benchmark sink: the comment now describes the benchmark's own allocation count, not how AgentPool.Refresh uses the result.

Left for follow-up PRs, still part of the leader-map fallback:

• Partition isolation. A topic Metadata has never returned, a topic-level error, an empty agent pool, or a Leader below 0 still makes ProduceSync fail the whole batch. The next PR routes the partitions that have a candidate and fails only the ones that do not.
• Stale leader still in the map. The named leader is present but the agent is gone. A later PR refreshes on unknown broker, dial failure, and a short list of topology errors. It does not retry the write already in flight. Timeouts and connection resets stay out.