Use io_uring to avoid blocking a thread while copying

[vlourme]

After 2.5-5k concurrent connections, the program tends to reach the 10k goroutines limit causing more latency.
Using io_uring, we could prepare splice and close without having to block a thread.

[vlourme]

After some trials with rio lib, results are not satisfying. I can't get better performance and it's mostly unstable.

In fact, performance are already okay-ish, but there are two bottlenecks:

• One connection per thread design lead to too much thread being blocked/parked while waiting for transferring data
• Golang does not always splice data by default, but read then write, making kernel to user-space copy.

Re-writing the whole program to io_uring is to complicated, parsing and auth validation are already working nicely in Go and it does not seems to be the blocking part.

I learned about SCM_RIGHTS recently, basically, after validating a request and before establishing a connection to the upstream. We could use sengmsg and SCM_RIGHTS to delegate the downstream file descriptor to another program who will take care of the copying process.

I will likely use Zig for that, it comes with a great io_uring implementation. Via Unix socket, it will receive a "transfer" request with a payload containing the IP of the target and eventually a "pre-writing" step (for HTTPS support, yet to decide).
It will then queue multiple prep calls into io_uring and execute that.