[GPU] Make permuteWeights inline (pytorch#47634)

Summary:
Pull Request resolved: pytorch#47634

Follow up on d16r's diff - D24710102. Make the function inline in order to get rid of the compiler checking `-Werror,-Wunused-function`.
ghstack-source-id: 116607200

Test Plan:
1. Sandcastle Tests
2. CircleCI jobs

Reviewed By: d16r

Differential Revision: D24824637

fbshipit-source-id: c17e219b384b91ac4620aa23112a6cda1200a605

aten/src/ATen/native/metal/MetalPrepackOpContext.h

@@ -70,7 +70,7 @@ class Conv2dOpContext : public torch::jit::CustomClassHolder {
 
 // The MPSCNNConvolution class takes weights in the order
 // [outputChannels][kernelHeight][kernelWidth][inputChannels/groups].
-static std::vector<float> permuteWeights(
+static inline std::vector<float> permuteWeights(
     const float* src,
     const std::vector<int64_t>& sizes) {
   const int64_t M = sizes[0];