freedreno: drawoverhead opts

Dropping non-softpin support for a6xx (since there was no overlap with kernels supporting a6xx and later and kernels that didn't support softpin) is a small (~1.5%) win at drawoverhead. The last patch to optimize re-emit of same stateobj to a submit multiple times is a bigger win (bringing total closer to ~5% win), with bigger ~2x win for things like shader state changes which trigger re-emitting lots of stateobj.

I should probably get a copy of whatever scripts @anholt has for generating stats from a larger # of drawoverhead runs (since my #s just based on a couple runs each, and pasting results into a gsheets), but the improvements seem big enough to be statistically significant.