Patch against upstream onnxruntime/core/mlas/lib/mlasi.h
Base commit: 62f742f1aa0c3102745ed35e3d869eaee845b9ac (ORT v1.26.0)

Reroute the public-header include to the relative vendored path,
and make MlasGetMaximumThreadCount() return cv::getNumThreads()
when MLAS_OPENCV_THREADING is defined (so MLAS partitions according
to OpenCV's thread budget, not ORT's).

--- a/3rdparty/mlas/lib/mlasi.h
+++ b/3rdparty/mlas/lib/mlasi.h
@@ -35,7 +35,9 @@
 #endif
 #endif  // MLAS_NO_EXCEPTION
 
-#include "core/mlas/inc/mlas.h"
+// Vendored under 3rdparty/mlas/. The ORT path "core/mlas/inc/mlas.h" only
+// works when MLAS is part of the ORT source tree.
+#include "../inc/mlas.h"
 
 #if defined(_WIN32)
 #ifndef WIN32_LEAN_AND_MEAN
@@ -1675,13 +1677,23 @@
     );
 
 
+#if defined(MLAS_OPENCV_THREADING)
+// Defined in 3rdparty/mlas/threading_opencv.cpp. Returns
+// cv::getNumThreads(). Hidden behind a free function so this header doesn't
+// need to pull <opencv2/core/utility.hpp> into every MLAS translation unit.
+extern "C" int opencv_dnn_mlas_max_threads();
+#endif
+
 inline
 ptrdiff_t
 MlasGetMaximumThreadCount(
     MLAS_THREADPOOL* ThreadPool
     )
 {
-#if defined(BUILD_MLAS_NO_ONNXRUNTIME)
+#if defined(MLAS_OPENCV_THREADING)
+    MLAS_UNREFERENCED_PARAMETER(ThreadPool);
+    return static_cast<ptrdiff_t>(opencv_dnn_mlas_max_threads());
+#elif defined(BUILD_MLAS_NO_ONNXRUNTIME)
     MLAS_UNREFERENCED_PARAMETER(ThreadPool);
     return 1;
 #else