Support retries in MultiClusterWrapper

[gbanasiak]

The multi-cluster-wrapper custom runner unwraps the runner before calling it for every remote cluster. This strips the wrapper responsible for retries (src) from the runner making operation parameters such as retry-until-success ignored.

Another wrapper removed during the unwrap is MultiClientRunner (src, src). This wrapper assumes that ES client is specified through a {"<cluster-name>": <client>} dictionary, and for single-cluster runners (all built-in Rally runners) this wrapper extracts the client from default cluster, while for multi-cluster runners (custom ones), provides entire dictionary.

This PR tries to address it by resigning from the unwrap. For this to work ES client is passed in a dictionary with default key which assumes the underlying runner will be a single-cluster one.

[gbanasiak]

I couldn't make unit tests work locally for me without Python and package upgrades.

[gareth-ellis]

How come we jump to 3.12 instead of 3.9 ?

[gbanasiak]

This follows dev environment preparation from Rally's Makefile, where 3.12 is used as well. It aligns both.

[gareth-ellis]

Ah, I hadnt spotted we had switched there (We used to use 3.8, then when we upgraded to 3.12 we jumped straight to 3.12). I suppose from a testing perspective we still test on 3.9, so thats ok I guess

[gbanasiak]

Exactly, we test 3.9-3.12, but for dev purposes configure 3.12.

[gbanasiak]

Actually we cannot assume that, this requires a modification.

[gbanasiak]

I've generalized in e3f42ea. Today multi-cluster-wrapper is only used with built-in Rally runners which are exclusively single-cluster. But I can imagine a situation where someone wants to iterate through all the clusters providing a dictionary of clusters for the runner to choose from. Perhaps that's too much.