(Filing this bug after a conversation with :kvark about texture upload improvements:)

At the moment there is a copy between data that the RenderBackend generates and the PBO main-memory representation that is used for texture upload.
We should eliminate this copy.

The hard part of doing so is communicating a pointer from the renderer thread (which calls the OpenGL function to map a buffer) to the RenderBackend thread (which needs to know where to write the data). We can do this by having a number of buffers around, and by having a fallback path in case there is no existing buffer. This fallback path would do what's done at the moment: Create an allocation on the RenderBackend thread to write the data into, and then copy the data into a new buffer mapping on the Renderer thread.