MultiGPU benchmark

I created this for personal testing, but I ended up cleaning it up
enough to the point where it can reasonably be published.

Tests the upload/download performance in a multi-GPU scenario, including
between different VkDevices created on the same GPU. (Just pass the same
GPU name twice to test this)

To bypass host pointer importing, for debugging and in some situations
where doing so would lead to quickly exhausting 4G limit.

To help during debugging. These should hopefully be disabled by any sane
release distribution, anyway.

Don't add it to the memory import failure paths because those can be
triggered spuriously during probing.