Consumer: pull loop dispatches each fetch's per-partition records to
a goroutine without a between-fetch barrier (no partitionWg.Wait
between PollFetches), so franz-go's local buffer keeps draining
while one slow partition is in flight. Goroutines drain into a
shared uncommittedRecords slice under a mutex. Shutdown waits on
a single partitionWg.Wait so the final commitRecords captures every
in-flight record. AutoCommitEnabled uses MarkCommitRecords (lock-
internal in franz-go, safe from many goroutines).

Producer: sendTxMetaToKafka / sendTxMetaBatch now set the message
Key to the (first) tx hash. With franz-go's default
StickyKeyPartitioner this hashes onto a single partition
deterministically, distributing traffic evenly across partitions
(tx hashes are uniform) and keeping every record from one batch on
the same partition. Nil keys previously degraded the
StickyKeyPartitioner to a StickyPartitioner which bunched batches
until linger expired.

txmetaHandler: shards Kafka ADD/DELETE entries by hash byte across
256 worker goroutines with bounded per-shard queues. Preserves
per-key ordering. Switches from blocking-on-full to drop-on-full
once any partition reaches its tail (txmetaCaughtUp latch is
one-way).

Adds SetCacheMulti to the txMetaCacheOps interface and a
SetTxMetaCacheMulti server method for future batch-cache fan-out.

Surfaced as a follow-up split out of PR bsv-blockchain#828; reviewable on its own.

…own path

Two correctness gaps from the Copilot review:

1. Per-partition concurrency across consecutive fetches could lose
   records. PollFetches returns immediately, so a second batch from
   partition P could enter its own goroutine and Mark/append a later
   offset before the previous batch had finished — a subsequent
   commitRecords would advance the partition past records the earlier
   batch had not yet processed, and if that batch later failed on an
   earlier offset the broker never re-delivered them. Add a per-
   partition mutex (lazily allocated via sync.Map) so batches from the
   same partition serialise; different partitions stay parallel.

2. The ErrClientClosed / context.Canceled branch in the error-handling
   path returned directly without partitionWg.Wait + final commitRecords.
   Records processed after the last ticker commit were left
   uncommitted despite successful processing. Factor the cleanup into
   a shutdownDrain closure and call it from both shutdown paths.

Plus fix the inaccurate SetTxMetaCacheMulti doc comment: the method
exists for a future batched fan-out, but the current shard-worker
txmetaHandler applies entries one at a time via
SetTxMetaCacheFromBytes and does NOT call SetTxMetaCacheMulti.