[Bug]: JMSIO drops messages when autoscaling down.

[jlampek]

What happened?

Steps to Reproduce

1. Create a Dataflow job that uses JMSIO to read from an MQ queue.
2. Configure the Dataflow job to use stream engine with number of workers initially set to 5 and minimum to 1.
3. Use the DefaultAutoscaler so the workers will auto downgrade from 5 to 1 when Dataflow cannot determine the backlog.
4. Publish messages to the MQ topic.

Expected Behavior
Workers should scale down, no messages should be dropped.

Current Behavior

1. The dataflow job is consistently dropping transactions at the point in time the workers autoscale down.
2. GCP Dataflow logs do not indicate that the runner is downscaling.

Issue Priority

Priority: 3 (minor)

Issue Components

• Component: Python SDK
• Component: Java SDK
• Component: Go SDK
• Component: Typescript SDK
• Component: IO connector
• Component: Beam YAML
• Component: Beam examples
• Component: Beam playground
• Component: Beam katas
• Component: Website
• Component: Spark Runner
• Component: Flink Runner
• Component: Samza Runner
• Component: Twister2 Runner
• Component: Hazelcast Jet Runner
• Component: Google Cloud Dataflow Runner

[Abacn]

May or may not be related to #25945. I was suspecting using connection pool that closes only in tearDown could cause problem #25945 (comment). If the publish is cached and only get flushed on closing session then data loss might happen, as teardown is not guaranteed to be called.