When sending a tail-loss probe a PING frame can be added to a DATAGRAM frame

[flub]

When a tail-loss probe needs to be sent the current queue_probe algorithm adds a PING frame even if a DATAGRAM frame could also be sent. This wastes a single byte, potentially amplified by being the start of a GSO batch.

This isn't that bad, there are many cases in which the user datagram would not fit into the probe packet anyway. Still, it could be a nice small improvement.

See #2169 which tried to solve this before, but probably broke some edge cases: https://github.com/quinn-rs/quinn/pull/2169/files#r1987103861

See #2172 which reverts this part of #2169 again.