clarify stale nonce handling in xDS protocol spec

Title: Clarify stale nonce handling in xDS protocol spec

Description: The xDS protocol spec currently says the following:

The management server should not send a DiscoveryResponse for any DiscoveryRequest that has a stale nonce. A nonce becomes stale following a newer nonce being presented to Envoy in a DiscoveryResponse.

This means that if the server sends an update to the client at the same moment that the client sends a new request changing the resource_names list, the client's request will be ignored by the server. The expectation here seems to be that this will not actually cause a problem, because the server will see the updated resource_names list when the client ACKs the response that the server just sent.

This approach could cause a problem in certain edge cases. For example, if the server is sending updates at a fixed interval that happens to exactly match the RTT to the client, then every time the client sends an ACK, it will be considered stale by the time it arrives at the server. In this case, the server will never see the updated resource_names list sent by the client.

This is admittedly probably fairly unlikely, but it would be good to clarify the protocol somehow to avoid this.

When I talked with @htuch about this, he said that management servers should ideally make sure that they only have one outstanding response to a client at any given time. In other words, a management server should ideally not send a new response until it sees the ACK/NACK for the previous response. However, the protocol spec does not say anything about requiring this.

One possible way to deal with this would be to change the spec to say that management servers must either (a) have only one outstanding response to a client at any given time or (b) not ignore requests with stale nonces. In other words, if the management server is not doing something to ensure that most of the client's requests will be non-stale, it cannot ignore stale requests.