GSO-Friendly Packet Memory

[victorstewart]

this trivial patch enables the batching of packets out into a persistent buffer per connection, for cleaner more efficient GSO.

currently there is no way to relate the peer_ctx back to a connection inside the pmi_allocate callback, so the best you can do is a pool of packet buffers. This can lead to ~40+ unique packet buffers being passed via sendmsg into the kernel and just as many unique copies.

or one could construct a contiguous buffer of all packets for GSO within the ea_packets_out at the expense of all those copies.

and all is complicated further by using io_uring given the asynchronous sendmsg completions.

this patch relies on the assumption of sequence. That for a given connection, the sequence in which packet buffers are requested from pmi_allocate matches the order in which they are passed to ea_packets_out.

now all you need to do is group packets by connection in your ea_packets_out callback, then calculate the contiguous buffer they constitute into a new struct iov_vec and pass it into sendmsg.

accidentally deleted my branch lol. So here's the PR again, including all the changes we discussed.