Support reading from replicas on pipeline executions

[FranGM]

This is the smallest possible change that should still satisfy #400

I've been testing it with our own client code and it seems to work as expected.

Something to note about this implementation is that it won't necessarily stick to the same replica for all read operations within the same shard. There are two reasons for this:

• changing this behaviour would complicate the implementation and probably require making the connection pools pipeline-aware somehow as well, or provide a custom connection pool for this use case.
• I actually think the behaviour as implemented is preferable as it will distribute the load across replicas, which is the original goal of this change anyway.

[FranGM]

since I had to move READ_COMMANDS to make it available to the pipeline object I ended up converting it into a frozenset (stolen from #406) since it does speed up lookups by an order of magnitude 😮

[Grokzen]

@FranGM This looks awesome and a much smaller change to get the feature out quicker. Please just dig into it and really confirm it without a doubt that commands is sent out to the replica nodes and i will take this for a run myself and i hope i can get a minor release out within this work week as this should be much faster to verify and get out :)

[FranGM]

Thanks for the quick reply!

I can do better than that actually :) I have it running in production right now. Here's a graph showing CPU usage on primary and its three replicas before and after the change

[Grokzen]

Good enough for me :) I aim for this weekend at the latest to go through it and get out 2.1.3 with this included

[Grokzen]

@FranGM Tested it out locally and i also wrote a example script for the readonly pipeline part that i will commit up after this merge to make it more easily testable for anyone who want to try it out.