feat(transport): accept borrowed instead of owned input datagram

[mxinden]

Stacked on top of #2184.

Previously process_input (and the like) took a Datagram<Vec<u8>> as input. In other words, it required allocating each UDP datagram payload in a dedicated Vec<u8> before passing it to process_input.

With this patch, process_input accepts a Datagram<&[u8]>. In other words, it accepts a Datagram with a borrowed view into an existing buffer (&[u8]), e.g. a long lived receive buffer.

---

Extracted out of #2093.

Part of #1693.

Pull request stack:

1. feat: don't allocate in UDP recv path #2184
2. (this one) feat(transport): accept borrowed instead of owned input datagram #2187
3. feat(udp): return borrowed Datagram on receive #2188
4. feat(bin): don't allocate in UDP recv path #2189

[mxinden]

Thanks for the input @martinthomson.

I applied your suggestion. process_xxx functions now take d: Datagram<impl AsRef<[u8]>>.

Given that there is no more need to pass a Datagram by reference, as one would instead pass Datagram<&[u8]>, this effectively reverts #1543.

I am sorry for the large diff. @martinthomson let me know whether this is the direction you were suggesting above.

Commits are still split into changes to neqo-common, neqo-transport (and neqo-http3), neqo-udp and neqo-bin.

[martinthomson]

This looks right. There's a lot of noise, but that just makes me want to land it faster.

I'm not 100% happy with the size of Datagrams being paired with having them passed by value, but I think that is something the compiler will manage for us mostly. (That is, SocketAddr is pretty big, and we have two of those, plus a byte, plus a slice. Copying that isn't super expensive, but marking that Copy means that the compiler won't help us avoid work that isn't constructive.) That said, I don't see any patterns here that are particularly problematic.