Fix aarch64 build break (pytorch#2055)

Summary:
Pull Request resolved: pytorch#2055

The aarch64 CUDA builds use D46213158 to disable F14 intrinsics for
compilations driven by NVCC/CUDA, instead the typical workaround that x86
uses: D34439017.

However, it looks like there's some issue preventing NVCC from parsing
the `F14SetFallback.h` code.  It turns out that we likely never use this
code from `.cu` sources, so this diff just drops an umbrella header and
uses fine-grained `#include`s to avoid F14.

Reviewed By: meyering

Differential Revision: D49792747

fbshipit-source-id: 8d2ef8cc68bcb2442a5b34e521d548cbb03a4c09

fbgemm_gpu/include/fbgemm_gpu/permute_pooled_embedding_ops.h

@@ -9,7 +9,8 @@
 #pragma once
 
 #include <ATen/ATen.h>
-#include <torch/script.h>
+#include <torch/csrc/api/include/torch/types.h>
+#include <torch/csrc/autograd/custom_function.h>
 #include "fbgemm_gpu/ops_utils.h"
 #include "fbgemm_gpu/sparse_ops_utils.h"