Allow packets with impossible CIDs to be ignored rather than reset

[Ralith]

Reduces the likelihood of Quinn endpoints responding to non-QUIC packets.

I'm not sure whether this is needed, especially after #1794. However, the cost seems low, it's a more direct solution, and defense in depth is nice.

[djc]

Nice!

[djc]

I also wonder if the cost of crypto here actually makes it less attractive than the random generator? @lijunwangs do you think this would help your use case (on top of the other options we've merged recently)?

[Ralith]

That's a good question. A more thoughtful choice of hash function, rather than just using whatever ring has lying around, might help. Maybe even just std's hasher, since we're not trying to be super secure here...

[Ralith]

For clarity, extra cost is paid here in exactly two places:

• when generating a CID, which is gated behind the decision to respond to a connection attempt
• sending a stateless reset for an unrecognized packet

[Ralith]

Refactored to use a non-cryptographic hash.

[lijunwangs]

I also wonder if the cost of crypto here actually makes it less attractive than the random generator? @lijunwangs do you think this would help your use case (on top of the other options we've merged recently)?

Thanks @djc/ @Ralith -- this will be a great add to strengthen the security. Is the change backward compatible? Can it break existing client code?

[Ralith]

It should be backwards-compatible, yeah.