This is a very arguable change, feel free to reject. I'm seeing a consistent ~ 0.7 % send rate improvement with lower-end hardware at 10 GbE close to but not quite at max send rate.

• Change how batch is laid out in memory: Since all access is one packet at a time and never all IPs or all lengths, lay out all fields pertaining to the same packet adjacent in memory, for sequential access (single array of packets with metadata instead of separate arrays for packet data, lengths and ips)
• Reduce MAX_PACKET_SIZE from 4k to 2k minus metadata, which is still way above the standard MTU of 1500 that we probably don't want to exceed in order to avoid IP fragmentation

Extra benefits:

• Arguably cleaner code for accessing the packets in batches (no manual offset calculations)
• Single definition of MAX_PACKET_SIZE instead of the previous two definitions of MAX_PACKET_SIZE

Tested on the usual macOS Sonoma, FreeBSD 14 and Ubuntu 23.10 with and without netmap.