send_async / _handle_invite race

[err508]

Problem Definition

There is a race condition with send_async / _handle_invite. Triggered if:

• Client A invites
• The invite triggers _handle_invite in Client B's transport
• Client A starts sending messages to Client B
• Messages are lost, as the invite was not processed yet

Appeared during #3124 as seen here, related #2779, #3123.

[andrevmatos]

This is expected. Retry logic should handle this. How this was not the case?

[err508]

_handle_invite is (timewise) expensive for private rooms, the send_async retries go through before _handle_invite is completed.