[catpowder] Support memory-mapped rx and tx queues

[iyzhang]

Context

Linux packet capture for raw sockets is very inefficient. It only sends or receives a single packet at a time and requires a system call for every packet. This inefficiency is particularly bad for workloads that use small packets.

Proposed Solution

Implement support for memory mapped queues
We should support user-level send and receive queues. This lets us get benefits similar to io_uring but with our own faster TCP stack.