Update on "Reuse GELU implementation from PyTorch core"

kernels/optimized doesn't need to support embedded systems, so it can just take a header-only dep on PyTorch.

Note that, because we will pick up Sleef internally and ignore it
externally thanks to ATen vec, this PR gets to enable optimized GELU in OSS.

Testing: CI to make sure this doesn't break mobile build modes; happy to take advice on anything not currently covered that might break.

Differential Revision: [D66335522](https://our.internmc.facebook.com/intern/diff/D66335522/)

[ghstack-poisoned]

Author: swolchok

build/Utils.cmake

@@ -329,8 +329,12 @@ endfunction()
 macro(find_package_torch_headers)
   # We cannot simply use CMAKE_FIND_ROOT_PATH_BOTH, because that does
   # not propagate into TorchConfig.cmake.
-  set(OLD_CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ${CMAKE_FIND_ROOT_PATH_MODE_PACKAGE})
-  set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE BOTH)
+  foreach(mode_kind IN ITEMS PACKAGE LIBRARY INCLUDE)
+    set(OLD_CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} ${CMAKE_FIND_ROOT_PATH_MODE_${mode_kind}})
+    set(CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} BOTH)
+  endforeach()
   find_package(Torch CONFIG REQUIRED)
-  set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ${OLD_CMAKE_FIND_ROOT_PATH_MODE_PACKAGE})
+  foreach(mode_kind IN ITEMS PACKAGE LIBRARY INCLUDE)
+    set(CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} ${OLD_CMAKE_FIND_ROOT_PATH_MODE_${mode_kind}})
+  endforeach()
 endmacro()