Cache VAOs between runs.

Cache VAOs between runs.

NVIDIA's drivers (at least the Linux drivers) seem to have a bug
where, if you call glVertexAttribPointer() in one thread together
with some specific activity in another thread (at least another
glVertexAttribPointer() will do, but I've also seen it during
glXSwapBuffers()), it will try to free a non-malloced pointer,
crashing everything. This happens very rarely, but often enough
that it's a real problem for Nageru.

We solve it by simply pre-creating all needed VAOs ahead-of-time
and caching them, which makes us call glVertexAttribPointer()
a _lot_ less often (basically never, instead of several times
per frame). This might actually be a bit slower than the old code
(I haven't tested), but at least it works around the driver bug.

ABI break, unfortunately, but no API break.