This changes the drawing paradigm a bit. The VAO configuration is done
JIT-style and depends on context active shader.
This is to allow more flexibility for implementations to do optimization
at lower level.
The vao cache is now its own class to isolate the concept. It is this
class that is reference by the GLContext for ownership of the containing
VAO ids.
We instead use a handle reference counter on the GPUVertBufs used by
the instancing batches. This make sure that if an update happens on the
GPUVertBuf used to contruct the batch, they will never have the same
memory address than the previously allocated ones (since they are still
pending deletion thanks to the refcounter).
This avoid the linear search to update the GPUBatch in the case a
batch is deleted (which was even a bad option since they could be only
cleared)
