Fix test_vulkan_interop_image descriptor/command-buffer reuse hazard in useSingleImageKernel path

[shajder]

Fixes #2800

Environment: rusticl -> Zink -> Lavapipe (software), Mesa 26.3.0-devel, LLVM 17.0.6, VK_LAYER_KHRONOS_validation 1.4.313, Vulkan instance 1.4.358.

Fixes VUID-vkUpdateDescriptorSets-None-03047 and VUID-vkBeginCommandBuffer-commandBuffer-00049, only visible with options --useSingleImageKernel --useValidationLayers together.

vkDescriptorSet and vkCopyCommandBuffer were reused across the per-image loop in run_test_with_one_queue and run_test_with_two_queue without waiting for the previous iteration's dispatch to finish. Added a VulkanFence to synchronize reuse.

[rjodinchr]

Can you explain why we should use a VulkanFence and not just call waitIdle?
If we keep VulkanFence, I think we should use std::make_shared instead of new, and we could have something like that:

if (i2DIdx < vkImage2DList->size() - 1)
{
  vkQueue.submit(...);
  vkImageDispatchFence->wait();
  vkImageDispatchFence->reset();
}

[shajder]

Can you explain why we should use a VulkanFence and not just call waitIdle?

vkQueueWaitIdle drains the whole queue; from Khronos's own guidance (Vulkan-Samples wait_idle sample):

Don't: Use vkQueueWaitIdle() or vkDeviceWaitIdle() unless absolutely necessary for less granular synchronisation.
Do: Use Fences to asynchronously read back data to the CPU.

...

If we keep VulkanFence, I think we should use std::make_shared instead of new...

Done.