freedreno: Few more CPU overhead reductions

The main thing is freedreno: Re-work dirty-resource tracking which tries to avoid draw time resource tracking when the resource has already been tracked. This limits the tracking to first usage of the resource in a batch.

The ideal thing would be to do all this at state bind time, but (a) it would be difficult to retrofit the guarantee that fb state is always the first state bound, and (b) that all state is re-bound after the fb state is bound. But this gets us pretty close to the minimal necessary tracking.