Move GPU cache updates to the RenderBackend

[zakorgy]

The current solution:

• The B::Device and Heaps<B> are shared across the Renderer and RenderBackend threads.
• We create and store the active PMBuffer in the RenderBackend, and we write the updates in place instead of sendening them to the Renderer.
• The active PMBuffer's B::Device buffer handle is shared with the Renderer so Device can use it in the descriptor sets.
• If we have to recreate the PMBuffer (upon clear or resize), the old PMBuffer is sent to the Renderer and we store it until the frame(s) which still use it are in flight.
• Upon PMBuffer recreation a BufferMemorySlice struct is also sent to the Renderer which is used for handling DeferredResolves. This is a workaround until the main Device logic is in the Renderer thread.

~The last missing thing is transitioning the PMBuffers to their proper state. For this we either record the transition in the RenderBackend and send it to the Renderer thread or send additional info from the state of the PMBuffer and use it for the transitions in the Renderer.~

---

This change is 

[kvark]

Thank you, this is amazing work! It's unfortunate that both the device and the PMBuffer need to be shared between threads now. I hope it's only temporary, as at the end of the day we should be handling everything on the render backend side.