Define good heuristics to use `collect_mode: breadth_first`

[jpountz]

breadth_first does not work when sub aggregations require scores, which is why we never use it by default today. However, we recently added the ability to know whether an aggregator requires scores or not (for now it just assumes that if a script is used then scripts are needed but we can improve it in the future, for instance I think we could have the information with lucene-expressions at least).

This means that we could now use breadth_first by default, we just need to figure out a good heuristic of when this is going to make the terms agg faster and/or more memory-efficient.

[speedplane]

It would be fantastic if the breadth_first / depth_first choice could be handled on the ES side.

Short of that, it would be nice to have some insight into whether one should be using breadth_first instead of depth_first. Right now, it's hard to know which one you should use other than testing them both and seeing which is faster. For example, if a counter returned how many temporary buckets were created in order to build the aggregation, one could get some insight into the internals of the aggregation and better optimize it.