perf: optimize DrainBatchesForOneNode O(n) re-enqueue scan on leader migration

[thomhurst]

Context

PR #576 added a re-enqueue scan in DrainBatchesForOneNode (RecordAccumulator.cs) to handle the case where a partition leader migrates between Ready() and Drain(). When GetPartitionsForNode(nodeId) returns empty, all non-empty partition deques are re-enqueued to _readyPartitions.

Concern

The O(n) scan over _partitionDeques runs inside the send loop. While the partitions.Count == 0 condition is expected to be transient (leader migration), it can also fire for:

• Newly-created topics where metadata hasn't populated yet
• Stale metadata after broker disconnection

For producers routing to many topics/partitions, this linear scan could add latency.

Possible improvements

1. Narrow the condition: distinguish genuine leader migrations from stale metadata (e.g., check if the node was recently known to own partitions)
2. Track which partitions Ready() consumed per node: only re-enqueue those specific partitions instead of scanning all deques
3. Cap the scan: limit re-enqueue to partitions that were recently sealed (within the last N seconds)

Current behavior

The scan is harmless correctness-wise — duplicate enqueues are documented as safe, and the scan only fires on the rare partitions.Count == 0 path. This is a performance optimization, not a bug fix.

Introduced in PR #576, commit 09979d2.