freedreno: Reduce constbuf0 upload overhead (for >120 default uniforms)
Based on !4967 (merged), this is just the last two commits, with the final commit being the big one:
drawoverhead throughput (n=10 each)
-----------------------
no statechange: no difference
program statechange 19% +/- 5.17278%
8 UBOs change 3.65491% +/- 1.1642%
uniforms change (1/few uniforms) -13.1139% +/- 0.973516%
uniforms change (1/many uniforms) 102.031% +/- 22.8545%