Design: Events eviction and Dead Letter Queue

In the context of upcoming internal queuing persistence, we will introduce pipeline-wide reliability to make sure that queued events will not be removed until acknowledged as fully processed by the final pipeline stage. This will ensure at-least-once delivery semantic in particular in the case of an unsupervised stop or a crash.

In that context, events flowing in the pipeline will require to **always** be acknowledged in the normal runtime. The only situation where events are not acknowledged is upon an unsupervised stop or a crash. This requirement to always acknowledge events introduces the need to evict events upon processing errors or catched/uncatched exceptions.

These should be the possible outcomes of event processing:

```
1. the process is successful, the event/batch is ACK'ed
2. the process is blocked, ACK cannot happens until the process is unblocked
3. the process crashes with an uncatched exception, the event/batch is ACK'ed and evicted (DLQ)
4. the process catches an exception:
4.1. if there's a retry mechanism, the event/batch is retried
4.1.1. the retry is successful, the event/batch is ACK'ed
4.1.2. the retry fails (max retry?), the event/batch is ACK'ed and evicted (DLQ)
4.2. there's no retry mechanism, the event/batch is ACK'ed and evicted (DLQ)
```

We need to introduce a proper/holistic logstash-wide events eviction mechanism.