Unbreak torch.is_vulkan_available() on Mac (re-send of #154675, please stamp) (#155595)

This is a new PR duplicating #154675 due to merge issues with that PR coming from my old (now updated) version of ghstack.

I am a Vulkan noob, but this extension and flag seem to be necessary. See "Encounted VK_ERROR_INCOMPATIBLE_DRIVER" at https://vulkan-tutorial.com/Drawing_a_triangle/Setup/Instance .

(For anyone trying to repro at home, I have the following homebrew packages installed, not all of which may be necessary: molten-vk, vulkan-headers, vulkan-loader, vulkan-tools, vulkan-utility-libraries. I also have VK_ICD_FILENAMES set to /opt/homebrew/etc/vulkan/icd.d/MoltenVK_icd.json, and I built PyTorch with USE_VULKAN=1. Making sure vkcube works helped me debug this setup.)
Pull Request resolved: #155595
Approved by: https://github.com/malfet

Author: swolchok

aten/src/ATen/native/vulkan/api/Runtime.cpp

@@ -78,6 +78,9 @@ VkInstance create_instance(const RuntimeConfiguration& config) {
 #ifdef VK_EXT_debug_report
         VK_EXT_DEBUG_REPORT_EXTENSION_NAME,
 #endif /* VK_EXT_debug_report */
+#ifdef __APPLE__
+        VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME,
+#endif // __APPLE__
     };
 
     find_requested_layers_and_extensions(
@@ -90,7 +93,11 @@ VkInstance create_instance(const RuntimeConfiguration& config) {
   const VkInstanceCreateInfo instance_create_info{
       VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO, // sType
       nullptr, // pNext
+#ifdef __APPLE__
+      VK_INSTANCE_CREATE_ENUMERATE_PORTABILITY_BIT_KHR, // flags
+#else // __APPLE__
       0u, // flags
+#endif // __APPLE__
       &application_info, // pApplicationInfo
       static_cast<uint32_t>(enabled_layers.size()), // enabledLayerCount
       enabled_layers.data(), // ppEnabledLayerNames