RPC connection pool should be aware of application-level errors

Currently the RPC connection pool (in `agent/pool/` and some surrounding code) is only aware of connection errors.

If a server is accepting connections, but all RPC requests are failing, the connection pool will not be aware of these failures.
We would also like to introduce a new error that servers will send back when they are not yet ready to handle RPC requests (possibly because they are in the process of loading a snapshot).

https://github.com/hashicorp/consul/blob/v1.10.1/agent/pool/pool.go#L591-L621 will need to look at `reply` and if its an error decide if it should `p.clearConn` or not.

This may be related to #10739