Introduce group_replica partial response strategy

[hczhu-db]

This is to implement the idea in https://docs.google.com/document/d/1eGoUFNGq2pOhmQH_svBkUZ45ig8bVOZH6xi9wkNRN_E/edit#heading=h.axx9hx8o4mz5

This strategy takes effect only when the partial response tolerance flag is off. Therefore, by default, this strategy is off.

Tested the image in dev-aws-eu-west-1 without turning on the new strategy.

How does the new strategy work?

1. A repeatable command line flag endpoint specifies multiple DNS names, like dnssrv+db-rep0:9092, dnssrv+db-rep1:9092, and dnssrv+range-store:9092

endpoints := extkingpin.Addrs(cmd.Flag("endpoint", "Addresses of statically configured Thanos API servers (repeatable). The scheme may be prefixed with 'dns+' or 'dnssrv+' to detect Thanos API servers through respective DNS lookups.").
		PlaceHolder("<endpoint>"))

2. Querier resolves each DNS name and gets a set of endpoints. Each endpoint set has a replica key and a group key parsed from the DNS name according to a pre-defined format.

Group key: db
Replica key: db-rep0
Endpoints: IP0, IP1, ...
---
Group key: db
Replica key: db-rep1
Endpoints: IP0, IP1, ...
---
Group key: range-store
Replica key: rang-store
Endpoints: IP0, IP1, ...

3. Querier fans out a query to all endpoints, even if one is determined unhealthy (different from other strategies). Querier counts how many failures from each group and replica and determines if the query result is actually complete.

[jnyi]

LGTM, nice, just a few small comments, do we have any tests to see how this behaves? Also, please make sure e2e tests pass to avoid regressions

[hczhu-db]

LGTM, nice, just a few small comments, do we have any tests to see how this behaves? Also, please make sure e2e tests pass to avoid regressions

some end-to-end tests are flaky on db_main.