fix(consumer): avoid broker race in response feeder

[DCjanus]

Background

A previous CI run in PR #3419 exposed a -race failure in Unit Testing with Go oldstable caused by concurrent access to partitionConsumer.broker.

Root Cause

responseFeeder() reads child.broker, but dispatcher() can change the same field during abort or redispatch. That is the race reported by -race.

The response path does not really need the current value of child.broker. It only needs to know which brokerConsumer produced the response that is being handled.

Fix Approach

Pass that brokerConsumer together with the feeder response.

This removes the read of child.broker from responseFeeder() and keeps the ack / resubscribe path tied to the broker that produced the in-flight response.

Testing

This PR was validated in three commits.

The first commit only adds the regression test plus a temporary workflow that runs just TestPartitionConsumerBrokerRace under -race, and is expected to fail. The second commit only contains the fix and is expected to make that workflow pass. The third commit removes the temporary workflow after the failure and recovery have both been demonstrated.

This sequence is intended to prove two things explicitly:

• the new regression test really detects the race in CI
• the follow-up fix really removes the race that the new test exercises

Relevant CI jobs:

• original motivating failure: Unit Testing with Go oldstable from PR #3419
• failing job: Consumer Race Repro
• passing job: Consumer Race Repro