[Tracking]: Performance issue of PubSubIO on non-Dataflow runner

### What happened?

This issue is used to track Beam PubSubIO performance issue.

Per https://cloud.google.com/dataflow/docs/concepts/streaming-with-cloud-pubsub Dataflow runner uses an internal implementation of PubSubIO. While the Beam's open-sourced PubsubIO is used for Direct Runner, Flink Runner, etc

It is observed that the Beam's implementation is less performant than the Dataflow one, which is as expected. However, the generic expectation is that the performance to be reasonable on other runners and we expect the SDK to be in production grade, which appears not the cases currently.

For example, there are reports of high ack and sent message counts when read from PubSub. It's not the high throughput use case. It's only 1 message per second. Yet, the number of "acks" is 14 times the number of published messages. And the number of "sent" is 6 times the number of published messages.

As a first step we should investigate why the messages are ack and published multiple times on, for example, the Flink runner.

### Issue Priority

Priority: 2 (default / most bugs should be filed as P2)

### Issue Components

- [ ] Component: Python SDK
- [X] Component: Java SDK
- [ ] Component: Go SDK
- [ ] Component: Typescript SDK
- [X] 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