Freedreno: Extend lower_io_offsets pass to lower image store/atomics on a4xx/a5xx
This is an improved rework of a series sent to mesa-dev a few months back.
It moves to NIR the offset computation of image store and atomics currently happening in the backend during NIR -> IR3 compilation.
It only benefits a4xx and a5xx because newer HW uses a different instruction and we don't need to emit code to calculate the offset.