Use `last_over_time` if no time-series agg is provided

[kkrik-es]

Since data points in each time series may be sampled independently, they are likely not aligned to the bucketing intervals of aggregation queries. Prometheus uses the newest sample (last value) that is less than the lookback period ago (5 minutes by default, query.lookback-delta flag) in each bucketing timespan, unless the query specifies a per-time-series aggregate function that is evaluated over all the values in the bucketing timespan. The plan is to introduce a similar approach for this type of queries.

This will unblock supporting queries missing a per-time-series inner aggregation function, like:

TS metrics | STATS avg(cpu_usage)) BY TBUCKET(1 hour)

[elasticsearchmachine]

Pinging @elastic/es-storage-engine (Team:StorageEngine)

[jdcryans]

@dnhatn chatting with @kkrik-es, I understand that you're looking at this so I assigned it to you.

[kkrik-es]

#134260 covers the use of last_over_time. We're missing the staleness check for PromQL, but this can be added later for that language specifically.