data: complete SGEMM upstream from 3 repos (siboehm+wangzyon+edtallison) + xllm fused_qknorm_rope + xattention kernels

SGEMM repos (upstream_ref/sgemm_cuda/, 41 files):
  siboehm/SGEMM_CUDA: kernel 1-12, runner, CMake, cuBLAS benchmark
  wangzyon/NVIDIA_SGEMM_PRACTICE: kernel 1-7 (Chinese comments), utils
  edtallison/sgemm-cuda: kernel 01-09 (learning notes), Makefile

xllm kernels (ex_engine/xllm_kernels/cuda/):
  fused_qknorm_rope.cu + bind — saves 128 kernel launches/fwd
  xattention/ — 6 files from upstream xllm
  headers: corex_compat_utils.h, topk_last_dim.cuh
  ilu/CMakeLists.txt

SO_BUILD_MANIFEST.md — complete .so inventory and call chain analysis

ex_engine/xllm_kernels/cuda/headers/corex_compat_utils.h

@@ -0,0 +1,37 @@
+// corex_compat_utils.h — Lightweight replacement for xllm's utils.h
+// Removes glog/tvm dependencies for BI-V100 corex compilation
+// Provides CHECK macro via TORCH_CHECK and DISPATCH macros from device_utils.cuh
+
+#pragma once
+
+#include <torch/torch.h>
+#include <c10/cuda/CUDAGuard.h>
+
+// Replace glog CHECK with TORCH_CHECK
+#ifndef CHECK
+#define CHECK(cond) TORCH_CHECK(cond)
+#endif
+
+#ifndef CHECK_EQ
+#define CHECK_EQ(a, b) TORCH_CHECK((a) == (b))
+#endif
+
+#ifndef CHECK_GE
+#define CHECK_GE(a, b) TORCH_CHECK((a) >= (b))
+#endif
+
+// Include device_utils for DISPATCH_HALF_TYPES etc
+#include "device_utils.cuh"
+
+// ffi namespace stub (some headers reference it)
+namespace ffi {
+template <typename T>
+using Array = std::vector<T>;
+}
+
+// HOST_DEVICE_INLINE
+#if defined(__CUDACC__) || defined(_NVHPC_CUDA)
+#define HOST_DEVICE_INLINE __host__ __device__ __forceinline__
+#else
+#define HOST_DEVICE_INLINE inline
+#endif