Outgoing datagram buffer accounting can panic the sender in quinn-proto

The limit on memory used to buffer outgoing datagrams, introduced in 0.11.17, subtracted an evicted datagram's size from the buffered byte counter twice when Datagrams::send was called with drop = true (as Connection::send_datagram in the quinn crate does) while the buffer was full. Under sustained eviction the counter drifts below its true value: debug builds panic with an integer underflow, while release builds wrap the counter, drain the whole queue, and then panic on a desynchronization check in a later send call. In quinn, this panic occurs while the connection lock is held, poisoning it and causing a double panic (and thus an abort) when the connection is dropped.

This can be triggered without a malicious peer, by an application that sends a burst of datagrams and then idles. It can also be triggered by a peer that keeps the connection alive with PING frames while withholding acknowledgements, so that the sender's congestion window stays closed and the outgoing queue never drains. Applications that use send_datagram_wait are not affected.

Only 0.11.17 is affected; the flaw was corrected in 0.11.18 by removing the redundant subtraction.

Thanks to @alexpyattaev for reporting and @vcarus for fixing this issue.