Query profiles parquet data from Queriers

As of now when we query from profiles we ask ingesters and store-gateways for profiling data. Most of data selection and aggregations/merge happens on those two components. This is where most of the CPU usage is on the read path.

Store-gateways are not really scalable since only 3 replicas of store-gateways has one given block per tenant, it also means some gateway will be more challenged than others which can cause unbalance usage ressources.

This is why I propose we stop selecting and merging data from store-gateways. We should only requests SeriesIndex+Labels per block to store-gateway then we should query directly parquet files from queriers and merge the data there. In fact I think store-gateways should only provide cache for tsdb and parquet file index.

This will reduce cost of operations since store-gateways have disk and will be less challenged but also improve scalability of the read path.

Start by tracing profile reads across ingesters, store-gateways, and queriers, focusing on SeriesIndex, labels, and parquet data selection and merging. Done means profile queries no longer select and merge data in store-gateways, while queriers query and merge the parquet data with the required behavior preserved.

Written by the indexing model from the issue text.