feat(windows): gfx1102 CPU-MoE serve stack (hybrid, 8 GB)

[AronVanAmmers]

Summary

Bundled Windows/ROCm changes from the gfx1102 report on #82, ported onto this fork (not a dump of FlashML PR FlashML-org#132).

On a Framework Laptop 16 (RX 7700S, gfx1102, 8 GB, 96 GB RAM with 8.2 GB hardware reserved) openai/gpt-oss-20b served at ~10.5 tok/s with CPU experts and PCIe fetch off. Offload / default hybrid stay at ~0.02 tok/s; --moe-hybrid-max-fetch 0 is a bypass, not a HIP gather fix.

This fork already has TCP ZMQ, a global Selector policy, and GPU offload TDR work. This PR adds what was still missing for the 8 GB CPU-MoE path:

• Build _cpu_moe on Windows with clang-cl + compiler-rt (MSVC would be scalar-only). Hoist <windows.h> out of the anonymous namespace so clang-cl atomics compile. Include hip_compat.cuh on ROCm.
• physical_core_cpus() via GetLogicalProcessorInformationEx (sysfs fallback counted SMT as cores: 16 threads on 8, ~3x less CPU-MoE bandwidth).
• Skip init_process_group when TheRock torch is USE_DISTRIBUTED=0 at TP size 1; guard all_reduce / shutdown.
• Shared zmq_addr(): zmq.has("ipc"), crc32, ports below 49152, per-instance suffix (the previous SHA1 ports could land in the Windows ephemeral range and did not isolate two servers).
• uvicorn loop="asyncio:SelectorEventLoop" (policy alone is not enough on uvicorn 0.52). Floor uvicorn>=0.36.

tvm-ffi hipcc flags are still a separate overlay: https://github.com/AronVanAmmers/tvm-ffi/tree/windows-hip-0.1.13-post3 (TVM_FFI_ROCM_ARCH_LIST=gfx1102).

Test plan

• pytest tests/utils/test_zmq_addr.py
• Framework 16 / gfx1102, clang-cl, TheRock HIP, overlay tvm-ffi, then:

$env:HIP_VISIBLE_DEVICES = "0"
$env:TVM_FFI_ROCM_ARCH_LIST = "gfx1102"
python -m freetoken.cli serve --model C:\models\gpt-oss-20b `
  --moe-backend hybrid --moe-cpu-threads 8 --moe-hybrid-max-fetch 0 `
  --moe-cache-auto --host 127.0.0.1 --port 1919

• One-token completion (not /health) as the readiness gate