Investigate performance impact of vertex array and framebuffer rebinding

[leod]

Currently, for convenience, we just have a single global VAO and a single global framebuffer whose attachments are updated in each draw call. We need to check if this has performance implications.

Note that we do not want to expose these objects to the user-facing API, since having to maintain them makes the API more cumbersome. However, we could consider to have an internal caching layer. IIRC, glium does this for vertex arrays.