Make WIDTH_PACKED the default packing scheme (#2518)

Summary:
Pull Request resolved: #2518

## Context

Switch to use `WIDTH_PACKING` by default in preparation for ExecuTorch alpha release. The demo model is Stories 110M which uses matrix multiplication heavily, and the tensor shapes used in the model matches `WIDTH_PACKING` the most.

Will put out a PR later on to expose controlling the default packing via serialization.

ghstack-source-id: 219281502
exported-using-ghexport

Reviewed By: jorgep31415

Differential Revision: D55074635

fbshipit-source-id: ab9d6d9946cef4cf7a27028178312d50de0ecbbe

Author: SS-JIA

backends/vulkan/runtime/graph/GraphConfig.cpp

@@ -55,10 +55,11 @@ GraphConfig::GraphConfig() {
   enableStorageTypeOverride = true;
   storageTypeOverride = api::StorageType::TEXTURE_3D;
 
-  // For now, force TENSOR_CHANNELS_PACKED memory layout by default as we are
-  // still developing support for other memory layouts.
+  // For now, force TENSOR_WIDTH_PACKED memory layout by default as we are still
+  // developing support for other memory layouts. In the future memory layout
+  // settings will be serialized as part of the graph.
   enableMemoryLayoutOverride = true;
-  memoryLayoutOverride = api::GPUMemoryLayout::TENSOR_CHANNELS_PACKED;
+  memoryLayoutOverride = api::GPUMemoryLayout::TENSOR_WIDTH_PACKED;
 }
 
 void GraphConfig::setStorageTypeOverride(api::StorageType storage_type) {