Use labels.ScratchBuilder to produce aggregate series

[fpetkovski]

This commit updates aggregate operators to use the newly introduced labels.ScratchBuilder from Prometheus when calculating output series.

Showing only differences in the memory benchmark. It looks like there's a measurable difference for aggregates with grouping labels.

name                                                  old alloc/op   new alloc/op   delta
RangeQuery/sum_by_pod-8                                 18.7MB ± 0%    18.1MB ± 0%  -3.39%  (p=0.008 n=5+5)
RangeQuery/topk-8                                       9.61MB ± 0%    9.59MB ± 0%  -0.19%  (p=0.032 n=5+5)
RangeQuery/rate-8                                       30.4MB ± 0%    30.4MB ± 0%  +0.04%  (p=0.032 n=5+5)
RangeQuery/sum_by_rate-8                                20.1MB ± 0%    19.5MB ± 0%  -2.94%  (p=0.008 n=5+5)
RangeQuery/quantile_with_variable_parameter-8           34.0MB ± 0%    33.3MB ± 0%  -1.90%  (p=0.008 n=5+5)
RangeQuery/at_modifier_-8                               23.9MB ± 0%    23.9MB ± 0%  -0.01%  (p=0.016 n=5+5)
RangeQuery/at_modifier_with_positive_offset_vector-8    23.7MB ± 0%    23.7MB ± 0%  -0.01%  (p=0.008 n=5+5)
RangeQuery/complex_func_query-8                         31.7MB ± 0%    31.8MB ± 0%  +0.24%  (p=0.008 n=5+5)
RangeQuery/sort-8                                       28.3MB ± 0%    28.2MB ± 0%  -0.07%  (p=0.032 n=5+5)

It also adjusts all operators to make sure they set the pool buffer size before allocating output samples.

[fpetkovski]

I caught the cases where we don't set output buffer sizes by panicking in the pool if the size is 0. Maybe we can have an implementation of the pool which does this and use it in tests?

[MichaHoffmann]

lgtm

[saswatamcode]

LGTM!

[GiedriusS]

671d55a#commitcomment-122044835 seems like this degraded performance? 🤔 actually, just one run's CPU usage jumped so perhaps this was a flake

[fpetkovski]

Yeah I also saw that, it's pretty cool that the CI job can report this.

I will take a look at the change again to see if is legit.

[MichaHoffmann]

Whaaa thast pretty big regression for absent