Release with librdkafka >= 2.15.0 — KIP-848 (group.protocol=consumer) consumer fixes needed

## Summary

The latest node-rdkafka release (v3.6.1) still bundles **librdkafka 2.12.0**. librdkafka **2.15.0** (released 2026-06-30) contains important fixes for the KIP-848 next-generation consumer group protocol (`group.protocol=consumer`), which is GA since 2.12.0. We are running `group.protocol=consumer` in production through node-rdkafka and have hit consumer-reliability problems that match the bugs fixed in 2.15.0. Since node-rdkafka pins the bundled librdkafka, there is currently no npm-installable way to get these fixes.

**Request: publish a release bundling librdkafka 2.15.0 (or newer).**

## Environment

- node-rdkafka 3.6.0 / 3.6.1 (librdkafka 2.12.0)
- Brokers: Apache Kafka 4.0.0 (KRaft, Strimzi)
- Consumer config: `group.protocol=consumer`, `enable.auto.commit=false` + `commitSync()`, `auto.offset.reset=largest`
- Multiple consumer pods per group, autoscaled (frequent rebalances)

## What we observed with 2.12.0 + KIP-848

1. **Silent consumption stalls**: consumers stop fetching entirely for 1–3 minutes during group reconciliations, with no `event.error` emitted and nothing in logs, then resume.
2. **Frequent commit failures** with `Broker: The member epoch is stale` (errno 113) whenever members join/leave (autoscaling) — hundreds per day across groups.
3. **In rare cases, silently skipped offset ranges**: after reconciliation storms (e.g. a staggered `subscribe()` across all group members), a group's fetch position ended up past several thousand messages that were never delivered to the application. The messages were intact and consumable — other consumer groups on the same topic received all of them.

## Matching fixes in librdkafka 2.15.0

From the 2.15.0 changelog (all KIP-848 consumer fixes, #5541):

> - Defer the leave heartbeat until the assignment is revoked, so a member that exceeds `max.poll.interval.ms` rejoins cleanly instead of being rejected with a fatal `INVALID_REQUEST`.
> - Treat `GROUP_ID_NOT_FOUND` in the ConsumerGroupHeartbeat response as fatal, except while the member is already leaving.
> - Surface unexpected permanent broker-level heartbeat errors instead of retrying them in a loop.

The third item in particular matches the silent stalls we measured. librdkafka 2.14.2 also contains a relevant fix (SIGSEGV in `rd_kafka_cgrp_handle_LeaveGroup()` when the coordinator is unavailable during consumer close).

## Related (can file separately)

While debugging this we noticed `KafkaConsumerConsumeNum::Execute` (src/workers.cc) discards consumer error events — including `ERR__AUTO_OFFSET_RESET` — when they arrive after at least one regular message in the batch (`default:` case only calls `SetErrorBaton` when `m_messages.size() == eof_event_count`). This means offset-reset notifications never reach the JS layer, which made the skipped-offsets case above very hard to diagnose. Happy to open a separate issue for that if useful.

Thanks for maintaining this library!

Start by locating the configuration and release entry points that bundle librdkafka for node-rdkafka 3.6.1; verify how the bundled version is selected and packaged for npm. Update the bundled dependency to librdkafka 2.15.0 or newer, validate the consumer fixes against the stated KIP-848 configuration, and publish an installable release.

Written by the indexing model from the issue text.