Introduce ActionListener#empty()

[DaveCTurner]

Today we write ActionListener.wrap(() -> {}) for an empty, do-nothing
ActionListener in a number of places. This auto-formats badly. As suggested
here[1] this commit introduces ActionListener.empty() for this situation
instead.

[1] https://github.com/elastic/elasticsearch/pull/39629/files#r262021809

[elasticmachine]

Pinging @elastic/es-core-infra

[original-brownbear]

@ywelsch pinging you here since you forbid me to do exactly this the other day :)

[ywelsch]

@original-brownbear perhaps explain why I wanted to avoid introducing this. My reasoning is that we should avoid using empty listeners in production code as much as possible. The issue with empty listeners is that they provide very little insight if anything goes wrong. Most times, a non-empty listener that does just trace logging is better than having a noop listener. @DaveCTurner what do you think about introducing this for tests only (e.g. in test:framework)?

[DaveCTurner]

Indeed, today there is only one usage of this in production code, in IndicesClusterStateService, and it seems ok there because all calls to its failure handler are preceded by a logger.warn() call.

The usages in #39629 will be similar: there we could accumulate a number of listeners all waiting for the same connection/disconnection to complete, and it makes more sense to log the failure before notifying the listeners rather than within each listener.

However I see that this could encourage trappy behaviour - I had to put some effort into being sure that we do log every failure in that PR, and there is a risk if we introduce this API that its use will become more widespread and we will face hard-to-debug failures in future. On reflection, let's not do this.