GS/Vulkan: Use dynamic rendering instead of render passes

[stenzek]

Description of Changes

Since VK_KHR_dynamic_rendering_local_read is a thing now, we can use dynamic rendering and still have our feedback loops.

I'm mainly PRing this since it's a different approach that may help with the RDNA3 crashes. I don't have a RDNA3 GPU, so I have no idea if it does or not.

If it does help with RDNA3, then I'll re-add a render pass fallback path for Kepler and Polaris, since it doesn't support the local read extension, and we'd lose Vulkan support for those GPUs otherwise. But if it doesn't, it's potentially not worth bothering with this change in the first place.

Rationale behind Changes

~10% lower CPU usage on GS thread in barrier-heavy situations.

Suggested Testing Steps

Check RDNA3.