delay fast field type conversion to harvest phase

[PSeitz]

extract_typed_sort_value_opt is in the collect hot loop and very performance sensitive. Delay conversion from u64 to actual type to harvest phase.

The impact on the histogram query time seems a little odd, probably related to some inlining. The impact on other sort query times roughly what I'd expect.

[PSeitz]

A sad bit is that we compare things twice:

* the search_after part is acting as a first fence.

* the head of the top k collector is also acting as a fence.

In reality... there is only one fence of course. We could have had single set of sort values on the segment collector (current_score_fence) or something like that, initialized it on the search_after parameter, and then just used the regular collect_top_k logic.

What makes everything so difficult is the different treatment of ties ( the split_search_after_order thing).

True, I think we could maybe have specialized SegmentCollectors with Box<SegmentCollector>, to provide specialized implementations via generics for the different use cases. The calling performance overhead in combination with collect_block should be fine.

[fmassot]

should we merge this PR and open a follow-up issue then?

[fulmicoton]

should we merge this PR and open a follow-up issue then?

Let's fix the bug and then merge. People do rely on edge.