fix: Propagate headers/warnings/stats from quantile downstreams

[benclive]

What this PR does / why we need it:
Passes headers from Downstreams back to the caller

• This caused a bug where the cache header was injected in downstreams after sharding, but wasn't passed back up to the query frontend, so the results were never cached.
• I copied the implementation from the StreamAccumulator in the same file - so this also passes back warnings & statistics if they are present.
• I added a test to confirm.

While debugging, I added a Span for when the CacheGenHeader is added, and you can see the span is triggered, and the header injected, in the Downstream call but never passed all the way up to the ResultsCache.Do because of this bug:

Trace from my local instance showing that 3/4 intervals now hit the cache:

[jeschkies]

I see now what I did wrong. The stats, warnings etc are joined in Downstream here but my accumulator is returning only one result.

I don't know how adventurous your are but the current state has a little code smell. I think the join of the stats should be moved out of Downstream. Furthermore, if you follow the code for the streams you'll see that we run iter.NewSortEntryIterator on the results. However, NewStreamAccumulator(params) returns only one result. So the sort iterator is not really required.

I suggest to change the Accumulator.Result return type to logqlmodel.Result instead of a slice. And then move the logic of the merge for the BufferedAccumulator. I'd do this as a follow up. The big work is that this would require moving the logic of NewMergeLastOverTimeStepEvaluator, NewMergeFirstOverTimeStepEvaluator and NewConcatStepEvaluator into new Accumulators.

Stepping back a little, my refactoring from #11863 was intended to go down that path.

Let me explain why this makes sense:

1. It will enable us to move the joining of stats into a central logic so it's not lost on further refactoring.
2. Accumulate and drop samples as they arrive. E.g. last_over_time does not have to buffer all the results anymore. Same with sum by in the future. This should reduce the memory footprint a little.
3. This will allow us to define special topk or sum by accumulators in the future..
4. The previous point will then enable us to print and investigate a query plan without running the step evaluators.

TLDR:

1. We should land this and verify that it works.
2. Replace BufferedAccumulator with ConcatAccumulator, LastOverTimeAccumulator and FirstOverTimeAccumulator.