S390x simd implementation (#25757)

### Description
This change adds SIMD-optimized implementation of functions for s390x.
This implementation is based on similar functions for ppc64le.

#### Build System Integration (onnxruntime_mlas.cmake):
* Adds a new S390X flag to the CMake build system to detect the target
architecture.
* Includes new source files specific to s390x (SgemmKernel.cpp,
DgemmKernel.cpp, Quantize.cpp, qgemm_kernel_zvector.cpp, etc.).
* Sets the necessary compiler flags (-mvx, -mzvector, -march=z15) to
enable z/Vector extensions.

#### Platform Abstraction (mlasi.h, platform.cpp):
* Defines MLAS_TARGET_S390X and MLAS_ZVECTOR_INTRINSICS for conditional
compilation.
* Integrates the new s390x kernels into the MLAS_PLATFORM dispatch
table.
* platform.cpp now checks for z/Vector support at runtime using
getauxval(AT_HWCAP) and HWCAP_S390_VXE, allowing it to fall back to
scalar implementations if the hardware support is not present.

#### New Kernel Implementations:
* qgemm_kernel_zvector.cpp: Implements quantized integer matrix
multiplication. This is the core of the performance improvement for
quantized models.
* SgemmKernelZVECTOR.cpp / DgemmKernelZVECTOR.h: Implements single and
double-precision floating-point GEMM.
* QuantizeZVECTOR.cpp / Quantize.cpp: Implements quantization and
requantization kernels.
* FgemmKernelZVECTOR.h: A generic header providing templates and macros
for both single and double-precision GEMM, similar to the ppc64le
implementation.

### Motivation and Context
This change improves performance of onnxruntime on s390x.

Author: AlekseiNikiforovIBM

cmake/external/eigen.cmake

@@ -4,10 +4,13 @@ set(EIGEN_BUILD_LAPACK OFF CACHE BOOL "" FORCE)
 set(EIGEN_BUILD_PKGCONFIG OFF CACHE BOOL "" FORCE)
 set(EIGEN_BUILD_CMAKE_PACKAGE ON CACHE BOOL "" FORCE)
 
+set(PATCH_EIGEN_S390X ${PROJECT_SOURCE_DIR}/patches/eigen/s390x-build.patch)
+
 onnxruntime_fetchcontent_declare(
     Eigen3
     URL ${DEP_URL_eigen}
     URL_HASH SHA1=${DEP_SHA1_eigen}
+    PATCH_COMMAND ${Patch_EXECUTABLE} --binary --ignore-whitespace -p1 < ${PATCH_EIGEN_S390X}
     EXCLUDE_FROM_ALL
 )
 onnxruntime_fetchcontent_makeavailable(Eigen3)

cmake/onnxruntime_mlas.cmake

@@ -384,6 +384,8 @@ else()
           set(X86_64 TRUE)
         elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^loongarch64.*")
           set(LOONGARCH64 TRUE)
+        elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^s390x$")
+          set(S390X TRUE)
         endif()
     endif()
 
@@ -792,6 +794,24 @@ endif()
           set(MLAS_SOURCE_IS_NOT_SET 0)
         endif()
     endif()
+    if(S390X AND MLAS_SOURCE_IS_NOT_SET)
+        set(mlas_platform_srcs
+          ${MLAS_SRC_DIR}/s390x/SgemmKernel.cpp
+          ${MLAS_SRC_DIR}/s390x/SgemmKernelZVECTOR.cpp
+          ${MLAS_SRC_DIR}/dgemm.cpp
+          ${MLAS_SRC_DIR}/s390x/DgemmKernel.cpp
+          ${MLAS_SRC_DIR}/s390x/Quantize.cpp
+          ${MLAS_SRC_DIR}/s390x/QuantizeZVECTOR.cpp
+          ${MLAS_SRC_DIR}/s390x/qgemm_kernel_zvector.cpp
+        )
+        set_source_files_properties(${MLAS_SRC_DIR}/s390x/SgemmKernel.cpp PROPERTIES COMPILE_FLAGS "-DSINGLE")
+        set_source_files_properties(${MLAS_SRC_DIR}/s390x/SgemmKernelZVECTOR.cpp PROPERTIES COMPILE_FLAGS "-DSINGLE")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mvx -mzvector -march=z15")
+
+        if(NOT ONNXRUNTIME_MLAS_MULTI_ARCH)
+          set(MLAS_SOURCE_IS_NOT_SET 0)
+        endif()
+    endif()
     if(NOT ONNXRUNTIME_MLAS_MULTI_ARCH AND MLAS_SOURCE_IS_NOT_SET)
         file(GLOB_RECURSE mlas_platform_srcs
           "${MLAS_SRC_DIR}/scalar/*.cpp")