Network request backoff not coordinated across multiple threads

If network issues occur and connectivity is lost between the master and slaves, we retry network requests with exponential backoff. This is fine, but the backoff algorithm is not coordinated across threads -- each thread does its own backoff: all threads will retry after 0.1 seconds, then all will retry after 0.2 seconds, etc.

This occurred on a slave node but could affect the master as well.

No files or tests are named. Begin by locating the exponential-backoff retry path used for network requests between the master and slaves, then trace how concurrent threads schedule retries. Done means retries no longer synchronize into the same schedule across threads, with coverage for the affected slave-node behavior and any master path that shares it.

Written by the indexing model from the issue text.