Improve fallback logic on timeout

[efstajas]

If endpoint A times out once, we should use endpoint B for a while before trying A again, in order to avoid very slow RPC responses while endpoint A is temporarily degraded.