Update base for Update on "Remove ExecuTorch copy of Vectorized"

All uses are outside ExecuTorch core, so we can just use ATen Vectorized.

Differential Revision: [D66396016](https://our.internmc.facebook.com/intern/diff/D66396016/)

[ghstack-poisoned]

Author: swolchok

build/Utils.cmake

@@ -321,3 +321,20 @@ function(resolve_python_executable)
     )
   endif()
 endfunction()
+
+# find_package(Torch CONFIG REQUIRED) replacement for targets that
+# have a header-only Torch dependency. Because find_package sets
+# variables in the parent scope, we use a macro to preserve this
+# rather than maintaining our own list of those variables.
+macro(find_package_torch_headers)
+  # We cannot simply use CMAKE_FIND_ROOT_PATH_BOTH, because that does
+  # not propagate into TorchConfig.cmake.
+  foreach(mode_kind IN ITEMS PACKAGE LIBRARY INCLUDE)
+    set(OLD_CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} ${CMAKE_FIND_ROOT_PATH_MODE_${mode_kind}})
+    set(CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} BOTH)
+  endforeach()
+  find_package(Torch CONFIG REQUIRED)
+  foreach(mode_kind IN ITEMS PACKAGE LIBRARY INCLUDE)
+    set(CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} ${OLD_CMAKE_FIND_ROOT_PATH_MODE_${mode_kind}})
+  endforeach()
+endmacro()

kernels/optimized/CMakeLists.txt

@@ -63,14 +63,7 @@ message("Generated files ${gen_command_sources}")
 
 list(TRANSFORM _optimized_kernels__srcs PREPEND "${EXECUTORCH_ROOT}/")
 add_library(optimized_kernels ${_optimized_kernels__srcs})
-# We require Torch headers, which setup.py puts in CMAKE_PREFIX_PATH
-# for us. Toolchains that we might be using for cross-compiling could
-# set CMAKE_FIND_ROOT_PATH, which prevents find_package from finding
-# headers not rooted under CMAKE_FIND_ROOT_PATH. This is reasonable
-# for binary dependencies because they probably aren't built for the
-# target platform, but for our header-only use case, we should just
-# ignore CMAKE_FIND_ROOT_PATH.
-find_package(Torch CONFIG REQUIRED NO_CMAKE_FIND_ROOT_PATH)
+find_package_torch_headers()
 target_include_directories(optimized_kernels PRIVATE ${TORCH_INCLUDE_DIRS})
 target_link_libraries(
   optimized_kernels PRIVATE executorch_core cpublas extension_threadpool