Split UBO upload from non-UBO upload?

[anholt]

trafficstars

Right now we upload UBO and non-UBO uniforms at draw time when any state flag affecting the uniforms is set. This works out to being basically every draw call. Hoewver, UBO uniforms change a lot less frequently, only when the corresponding program or its constant buffer changes. We could potentially reduce our memory bandwidth, and draw-time CPU overhead by only re-uploading UBO contents when things affecting those change.

glmark2 and mupen64plus don't hit this. glbenchmark2.7 does.

[anholt]

vc4-ubo branch has an attempt at this. No effect on 3DMMES (which is FS limited, not CPU)