Update on "Dtype selective build: enable in fbcode"

^
By removing `header_namespace` and depending on #include "selected_op_variants.h" instead of the full path #include <executorch/kernels/portable/cpu/selected_op_variants.h>

Note:
expose_operator_symbols=False only works in xplat, so add a failure message for that too. I don't think we should recommend for users to set it to true, as it prevents a library from linking multiple executorch_generated_libs (symbols will clash).

Differential Revision: [D75082395](https://our.internmc.facebook.com/intern/diff/D75082395/)

[ghstack-poisoned]

Author: lucylq

shim_et/xplat/executorch/codegen/codegen.bzl

@@ -465,7 +465,7 @@ def build_portable_lib(name, oplist_header_name, portable_header_lib, feature =
     # library, and it blocks users like unit tests to use kernel
     # implementation directly. So we enable this for xplat only.
     compiler_flags = ["-Wno-missing-prototypes"]
-    if not expose_operator_symbols:
+    if not expose_operator_symbols and is_xplat():
         # Removing '-fvisibility=hidden' exposes operator symbols.
         # This allows operators to be called outside of the kernel registry.
         compiler_flags += ["-fvisibility=hidden"]
@@ -510,7 +510,7 @@ def build_optimized_lib(name, oplist_header_name, portable_header_lib, feature =
     # library, and it blocks users like unit tests to use kernel
     # implementation directly. So we enable this for xplat only.
     compiler_flags = ["-Wno-missing-prototypes", "-Wno-pass-failed","-Wno-global-constructors","-Wno-shadow",]
-    if not expose_operator_symbols:
+    if not expose_operator_symbols and is_xplat():
         # Removing '-fvisibility=hidden' exposes operator symbols.
         # This allows operators to be called outside of the kernel registry.
         compiler_flags += ["-fvisibility=hidden"]