[ET-VK] Fix staging buffer allocation to check all memory types for HOST_CACHED

`test_host_cached_available()` only checked `memoryTypes[0]` to determine if
HOST_CACHED memory was available. On Pixel devices, `memoryTypes[0]` is
DEVICE_LOCAL without HOST_CACHED, so the function incorrectly returned
`SEQUENTIAL_WRITE_BIT`. This caused DEVICE_TO_HOST staging buffers to be
allocated in write-combining (uncached) memory, making CPU reads during
COPY_OUTPUTS ~170x slower than necessary (~40ms vs ~237us on S24).

The fix iterates over all memory types to correctly detect HOST_CACHED support.

On-device profiling of edgetam_first_frame_fp16_vulkan.pte confirms the fix:
- Pixel 8 Pro COPY_OUTPUTS: 40ms -> 6.3ms (-84%)
- Pixel 9 Pro XL COPY_OUTPUTS: 40ms -> 2.5ms (-94%)
- Pixel 8 Pro Method::execute: 492ms -> 464ms (-5.7%)
- Pixel 9 Pro XL Method::execute: 445ms -> 411ms (-7.6%)

Differential Revision: [D97058156](https://our.internmc.facebook.com/intern/diff/D97058156/)

[ghstack-poisoned]

Author: ssjia

backends/vulkan/runtime/vk_api/memory/Allocator.cpp

@@ -16,13 +16,12 @@ VmaAllocationCreateFlags test_host_cached_available(
   VkPhysicalDeviceMemoryProperties mem_props;
   vkGetPhysicalDeviceMemoryProperties(physical_device, &mem_props);
 
-  VkMemoryPropertyFlags const flags = mem_props.memoryTypes->propertyFlags;
-
-  bool const host_visible = flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT;
-  bool const host_cached = flags & VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
-
-  if (host_visible && host_cached) {
-    return VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;
+  for (uint32_t i = 0; i < mem_props.memoryTypeCount; i++) {
+    VkMemoryPropertyFlags flags = mem_props.memoryTypes[i].propertyFlags;
+    if ((flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) &&
+        (flags & VK_MEMORY_PROPERTY_HOST_CACHED_BIT)) {
+      return VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;
+    }
   }
 
   return VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT;