Batched datagram I/O for Windows (IOCP/RIO) and io_uring on Linux

[Segfaultd]

Summary

Extend batched datagram I/O beyond the Linux recvmmsg/sendmmsg paths: io_uring on Linux as a next step, and IOCP with WSASendMsg/WSARecvMsg (or Registered I/O) on Windows.

Motivation

Linux currently coalesces up to MMSG_BATCH_MAX datagrams per syscall, which cut syscall counts by roughly 31x on burst workloads. Windows and macOS still pay one syscall per datagram, so servers on those platforms hit syscall overhead first under load.

Proposal

• Windows: batch sends through overlapped WSASendMsg submissions completed via IOCP, or Registered I/O (RIO) ring buffers for the server oriented build; RIO gives the closest analogue to sendmmsg amortization
• Linux (second phase): an io_uring backend submitting send/recv SQEs in bulk, replacing per tick sendmmsg calls with ring submissions
• Follow the established structure: all batching logic stays in portable helpers (see MmsgBatch.h: DriveBatchedSend, ClassifySendmmsgErrno equivalents, CompactRecvSlots) with unit tests compiled on every platform; only the syscall glue is platform gated, with no build options and no capability macros

Validation

• Unit tests for any new portable helper on all platforms
• On the target platform, a burst integration test that provably crosses the batch threshold (message counts above the batch max), mirroring Tests/Integration/MmsgBatchLiveTests.cpp
• Before/after syscall or completion counts recorded in the PR description using the platform profiler (strace equivalent, xperf, dtrace)

Acceptance criteria

• Behaviour identical to the per datagram path; only syscall counts differ
• No new configuration surface