To generate metrics based on observed errors that use the key and/or server, currently users need to do something like parse the message or override the check_return_code in order to hook into a place where they can attempt to observe the key used and infer the server involved similar to how the message is generated. This is additionally tricky around the ServerIsMarkedDead errors, since naively calling server_by_key would return the incorrect server (handled by detect_failure as part of message generation).

To ease users ability to track the keys and servers that errors are related to, this extends the base error class and attaches the same context applied to the messages to an attribute key_map.

As far as I can tell there's no differentiation between server-side and client-side errors, so the key_map's semantics are roughly "these were the keys used as part of this operation, and these are the servers we think would have been used with these keys"...a little wishy-washy but likely still useful to the user?