Expose more series size stats and use in lazy posting

[yeya24]

• I added CHANGELOG entry for this change.
• Change is not relevant to the end user.

Fixes #7955

Changes

• Add new stats field to meta.json index_stats section, including series_p90_size, series_p99_size, series_p999_size and series_p9999_size.
• Compactor will try to calculate p90 to p9999 quantile series size and write the stats to meta.json.
• Add new flag estimated-series-size-stat to configure which stats to use for lazy posting. By default it is max but users can configure their own to p90 to p9999. I didn't call this flag lazy-expanded-posting-estimated-series-size-stat because this estimated series size might be used by other features in the future for estimation.
• Lazy expanded posting uses the value above for series estimation.

Note:
For quantile series size calculation, I have tried both go-tdigest (already used by hedged requests feature) and DataDog/sketches-go. ddsketches-go seems 10s faster than go-tdigest when processing a block with 20M series. So I go with ddsketches-go in the implementation for performance.

Verification

Existing tests should still pass

[GiedriusS]

Same question as in the other PR: how does one know which choice is the "correct" one?

[yeya24]

We expose metrics about how much lazy posting optimized and how much we overfetch for series. If users want to optimize lazy posting more then they can use a smaller percentile stat.

It is hard to say correct or not. It is more like how to tune lazy posting to make it optimize more scenarios. In general, a block with 20M series, it is P9999 size should be only smaller than 2000 series. It is probably more like how much I am able to tolerate if things go wrong. And users can play with this config and the metrics above.

[yeya24]

Can you please take a look at this PR? @GiedriusS @MichaHoffmann