mesa,glsl: optimize _mesa_load_state_parameters for faster constant uploads
This reduces CPU overhead when there are too many constant buffer updates or too many state parameters (including e.g. local parameters for ARB programs).
Mainly ARB programs and fixed-function shaders benefit from this, like Viewperf.
I could do more than this, but I guess 17 commits is enough "for now".
Note: !8017 (merged) is a prerequisite for state variable merging.