feat(MoE): verified CUDA topk_softmax kernel — zero fallback

moe_topk_softmax_v3.cu: BI-V100 verified (2026-08-10)
  - 64 experts, topk=8, warp shuffle, zero shared memory
  - renormalize: sum=1.0 ✓, no NaN ✓, no duplicate ids ✓
  - 881 token batch ✓
  - Compiler: corex clang/16, --cuda-gpu-arch=ivcore10
  - Stream: c10::cuda::getCurrentCUDAStream()

corex_moe.py: loads CUDA kernel, NO Python fallback
  - Searches pre-compiled .so → JIT compile from source → error
  - MoE pipeline: CUDA topk → cublas expert GEMM → ixformer silu_and_mul

precompile_moe_topk.py: Docker build-time compilation + verification

Key finding from real machine probing:
  ixformer::infer::topk_softmax is DECLARED in ixformer.h but
  NOT IMPLEMENTED in any .so in the base image (nm -D scan: zero hits).
  Must compile our own kernel.

ex_engine/precompile_moe_topk.py

@@ -0,0 +1,34 @@
+"""
+Precompile moe_topk_softmax_v3.cu → .so during Docker build.
+Same pattern as precompile_gdn.py.
+
+Run: python3 ex_engine/precompile_moe_topk.py
+"""
+import os, sys
+
+def main():
+    cu_path = os.path.join(os.path.dirname(__file__), "csrc", "moe_topk_softmax_v3.cu")
+    if not os.path.isfile(cu_path):
+        print(f"[MOE] ERROR: {cu_path} not found")
+        sys.exit(1)
+
+    print(f"[MOE] Compiling {cu_path} ...")
+    from torch.utils.cpp_extension import load
+    ext = load(
+        name="moe_topk_softmax_v3",
+        sources=[cu_path],
+        extra_cuda_cflags=["-O3"],
+        verbose=True,
+    )
+    print("[MOE] ✓ moe_topk_softmax_v3.so compiled successfully")
+
+    # Verify
+    import torch
+    gating = torch.randn(4, 64, device='cuda', dtype=torch.float16)
+    w, ids, _ = ext.moe_topk_softmax(gating, 8, True)
+    assert not w.isnan().any(), "NaN in topk weights!"
+    assert torch.allclose(w.sum(dim=-1), torch.ones(4, device='cuda'), atol=1e-3)
+    print("[MOE] ✓ Runtime verification passed")
+
+if __name__ == "__main__":
+    main()