bypass4netns-style socket switching (opt-in host network mode)

[jserv]

Problem

All network I/O goes through the SLIRP/passt userspace stack. For bulk
transfers, data-path overhead dominates.

Proposed Changes

Add an opt-in --net=host-bypass fast path using SECCOMP_IOCTL_NOTIF_ADDFD
(the same mechanism already used for pipe injection in forward_pipe()):

1. Intercept socket() + connect() via seccomp for eligible sockets.
2. Create the socket in the host network namespace, inject via ADDFD.
3. Subsequent read/write/send/recv go directly to the host kernel
   without supervisor involvement on the data path.

Scope of first version:

• AF_INET/AF_INET6, SOCK_STREAM, outbound connect() only.
• Explicitly exclude AF_UNIX, AF_PACKET, raw sockets.
• Policy controls to restrict which destination addresses/ports are eligible
  for bypass.

Additional syscall interception beyond socket() is required:
connect, bind, listen, accept, setsockopt, getsockopt, and
fcntl on bypassed FDs all need consideration in the dispatch layer
(seccomp-dispatch.c).

Considerations

• epoll/poll multiplexing: LKL epoll cannot monitor host socket FDs and
  vice versa. Bridging both worlds may require intercepting epoll_ctl and
  epoll_wait to implement a unified event loop. This is the hardest
  architectural challenge and may limit applicability to simple
  connect-send-recv workloads initially.
• Network isolation tradeoff: the guest operates directly in the host
  network namespace for bypassed sockets. bind() binds to host interfaces,
  getsockname() returns host IPs. This must be explicitly opt-in with
  clear documentation of the security implications.
• Relationship to passt backend and SLIRP Phase 2 (SLIRP: server sockets and pidfd_getfd #12): this is a
  complementary approach. passt improves the stack-mediated path; socket
  switching bypasses the stack entirely for eligible connections.
• Address-family scope: start narrow (AF_INET/AF_INET6,
  SOCK_STREAM) and expand based on demand.