Instead of retrying safe reads 3 times, retry each eligible service once
Summary: Ref T13286. When retrying a read request, keep retrying as long as we have canididate services. Since we consume a service with each attempt, there's no real reason to abort early, and trying every service allows reads to always succeed even if (for example) 8 nodes of a 16-node cluster are dead because of a severed network link between datacenters.
Test Plan: Ran git pull in a clustered repository with an up node and a down node; saw retry count dynamically adjust to available node count.
Maniphest Tasks: T13286
Differential Revision: https://secure.phabricator.com/D20777