feat(consumer): Add commit log heartbeat for idle partitions

[yuvmen]

This is a draft for possibly fixing an issue in single tenants where we scaled the post-process-forwarder consumer from 1 to 16 partitions and are now seeing increased latency.

Summary

• When the events topic has low per-partition throughput (<1 msg/sec/partition across 16 partitions), some partitions may have no messages in a given batch window
• Snuba only produces commit log entries for partitions with data in the current batch
• The post-process-forwarder (SynchronizedConsumer) gates on the commit log — idle partitions get no entry, causing it to pause and wait, producing latency spikes
• This change emits a "heartbeat" commit log entry for all assigned idle partitions on each batch flush, re-using the last known committed offset
• Caps the maximum stall to one batch window (~500ms) instead of unbounded

Changes

• Python: CommitLogHeartbeatState shared across batch writer instances, heartbeat logic in ProcessedMessageBatchWriter.close() and MultistorageCollector.close()
• Rust: ProduceMessage tracks assigned partitions and last offsets, emits heartbeats for idle partitions
• Wiring: Heartbeat state flows through ConsumerBuilder → build_batch_writer() + KafkaConsumerStrategyFactory, updated on each rebalance via create_with_partitions()

Test plan

• 8 new Python unit tests covering: idle partition heartbeats, no heartbeat without prior offset, rebalance cleanup, offset evolution, no-state passthrough, MultistorageCollector, build_batch_writer shared state, strategy factory partition propagation
• 1 new Rust test: heartbeat entries produced for idle partitions with correct offsets
• All existing tests pass
• Deploy to staging and verify snuba-post-processor latency spikes are reduced

🤖 Generated with Claude Code