Streaming PromQL engine: sort series in matrix at end

[charleskorn]

What this PR does

This PR modifies the behaviour of the streaming PromQL engine to not sort series in the aggregation operator, and instead sort the final matrix value.

Matrix values are expected to be sorted by series labels, so we can't avoid sorting series altogether - but we don't need to sort intermediate results.

Sorting series in the aggregation operator is unnecessary:

• in the specific case of the aggregation operator, it's better to return series in whatever order minimises the number of incomplete groups held in memory, as this reduces peak memory utilisation in the case where the input series are sorted according to the output groups (and in the worst case, peak memory utilisation is no worse than what we have today)
• there may be another operator consuming the result of the aggregation, and so sorting the output may be a waste of time, as the consuming operator might not need its input to be sorted

Some operators (eg. selectors) will always return series in the correct order, but slices.SortFunc handles the case where the input is already sorted relatively efficiently, so this extra sorting pass does not add much overhead (<1% latency in our benchmarks).

The final sorting pass does not require more memory utilisation or allocations.

Which issue(s) this PR fixes or relates to

(none)

Checklist

• Tests updated.
• [n/a] Documentation added.
• CHANGELOG.md updated - the order of entries should be [CHANGE], [FEATURE], [ENHANCEMENT], [BUGFIX].
• [n/a] about-versioning.md updated with experimental features.

[pracucci]

It took me a while to understand why this change lower the peak memory utilization in some cases, but now I do understand it. Good job! 👏