Guard aten_device.h against -Wswitch-enum

The new header switches over c10::DeviceType to map a PyTorch device onto
an ExecuTorch one. That enum has over twenty values, and a build with
-Wswitch-enum turned on asks for every one of them to be listed even
though the switch already has a default. So the header does not compile
there.

This was hidden before. The same code used to live in aten_bridge.cpp, and
the aten_bridge library is built with -Wno-error, which demotes the
warning. Moving the code into a header exposed it to ETDump, which has no
such setting.

The fix wraps the file in
C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum"), the same way
extension/tensor/tensor_ptr.h already does for its own switch over a c10
enum. Plain -Wswitch is left alone, so forgetting to handle a new
ExecuTorch device type is still a compile error.

Three smaller corrections to the previous commit come with it:

- The comment on the new Buck target was wrong. Splitting device.h out
  does not keep the portable Tensor out of an ATen mode target, because
  the ETDump target already reaches it through
  runtime/core:device_allocator. The split is still worth doing, so that a
  caller wanting only Device does not depend on every portable type. That
  is the same reason :scalar_type is split out, and the comment now says
  so.
- aten_bridge.h no longer uses std::optional, so its <optional> include is
  removed.
- The new comments are shortened.

Test plan: compiled aten_device.h and the ETDump device helper with and
without USE_ATEN_LIB, under -Wall -Wextra -Wswitch-enum -Werror, against
PyTorch headers on Linux x86_64. All four are clean. Removing only the new
guard puts the 20 -Wswitch-enum errors back, so the guard is doing the
work. Adding a third ExecuTorch device type still fails the build through
plain -Wswitch, so the guard does not hide that. Also recompiled
aten_bridge.cpp and the exact call shapes the existing aten_bridge test
uses, in portable mode.

Author: shoumikhin

extension/aten_util/aten_bridge.h

@@ -8,10 +8,9 @@
 
 #pragma once
 
-// Portable mode only. The declarations below name torch::executor::ScalarType
+// Portable mode only: the declarations below name torch::executor::ScalarType
 // and torch::executor::Tensor, which exec_aten.h defines only when it is not
-// building against ATen. For the device conversions, which work either way,
-// include aten_device.h instead.
+// building against ATen. For device conversions either way, use aten_device.h.
 
 #include <executorch/extension/aten_util/aten_device.h>
 #include <executorch/extension/tensor/tensor.h>
@@ -23,7 +22,6 @@
 #include <c10/core/ScalarTypeToTypeMeta.h> // @manual=//caffe2/c10:c10
 
 #include <memory>
-#include <optional>
 #include <vector>
 
 namespace executorch {

extension/aten_util/aten_device.h

@@ -12,13 +12,16 @@
 #include <executorch/runtime/platform/assert.h>
 
 #include <c10/core/Device.h> // @manual=//caffe2/c10:c10
+#include <c10/macros/Macros.h> // @manual=//caffe2/c10:c10
 
 #include <optional>
 
-// These live apart from aten_bridge.h so that a translation unit compiled with
-// USE_ATEN_LIB can convert devices. aten_bridge.h is usable only in portable
-// mode: it names torch::executor::ScalarType and torch::executor::Tensor,
-// aliases that exec_aten.h defines only when it is NOT building against ATen.
+// Kept out of aten_bridge.h, which compiles in portable mode only, so that a
+// translation unit built with USE_ATEN_LIB can still convert devices.
+
+// c10::DeviceType has over twenty values and -Wswitch-enum wants every one
+// listed even with a default. -Wswitch still covers the switch below.
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum")
 
 namespace executorch {
 namespace extension {
@@ -65,3 +68,5 @@ torch_to_executorch_device(c10::Device device) {
 
 } // namespace extension
 } // namespace executorch
+
+C10_DIAGNOSTIC_POP()

extension/aten_util/targets.bzl

@@ -7,8 +7,8 @@ def define_common_targets():
     TARGETS and BUCK files that call this function.
     """
 
-    # Split out of :aten_bridge so that ATen-mode code can convert devices.
-    # :aten_bridge itself only compiles in portable mode.
+    # Split out of :aten_bridge, which compiles in portable mode only, so that
+    # ATen mode code can still convert devices.
     runtime.cxx_library(
         name = "aten_device",
         exported_headers = ["aten_device.h"],

runtime/core/portable_type/targets.bzl

@@ -46,9 +46,9 @@ def define_common_targets():
         ],
     )
 
-    # Device is the one portable type that ATen-mode code also needs, so it is
-    # its own library: depending on :portable_type from an ATen-mode target
-    # would drag in the portable Tensor alongside at::Tensor.
+    # device.h is standalone and is the one portable type header that ATen mode
+    # code also needs, so it is split out like :scalar_type below rather than
+    # making a caller that wants only Device depend on every portable type.
     runtime.cxx_library(
         name = "device",
         exported_headers = [