Split cpu/gpu in caffe2/distributed + some clean up (#20674)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20674

A few targets in caffe2/caffe2/distribute needs to be split too, otherwise won't compile. Also some clean ups and make select_gpu_type to gpu_library_selector

Differential Revision: D15406019

fbshipit-source-id: 6455ab885b248502b48d4c7565597e00fecfd547

tools/build_variables.py

@@ -3,7 +3,7 @@
 # not currently relevant so they are combined into one list.
 from __future__ import absolute_import, division, print_function, unicode_literals
 load("@bazel_skylib//lib:new_sets.bzl", "sets")
-load("//caffe2/caffe2/fb:defs_gpu.bzl", "gpu_library_targets")
+load("//caffe2/caffe2/fb:defs_gpu.bzl", "gpu_library_selector")
 
 GENERATED_CPP = [
     "Functions.cpp",
@@ -347,11 +347,11 @@ def add_torch_libs():
     )
 
     # TODO: split it into cpp and cuda parts similarly to libtorch
-    gpu_library_targets(
+    gpu_library_selector(
         name="_C_impl",
-        deps=[":_C_impl_cuda"],
         deps_cpu=[":_C_impl_cpu"],
-        merge_only=True,
+        deps_cuda=[":_C_impl_cuda"],
+        merge_cpu_deps=False,
     )
 
     cpp_library(