v3d: implement correct lowering for logic operations on MSAA render targets

@anholt: this mostly requires the ability to do sampled tlb reads and writes, and then it is a matter of emitting these for logic ops when the target is MSAA. For per-sample color writes I ended up adding another v3d-specific intrinsic mostly so I can differentiate these from normal color writes since we need to emit different code for them, and also because we also we need a sample index that I don't think we can get with standard store_output intrinsics (unless we figure out a creative way to pack that into the base index or something...).

I don't think there are any piglit tests that combine logic operations with MSAA render targets but I did hack the existing gl-1.0-logicop to work against a 4x RGBA MSAA render target and the test passes with this series, it might make sense to turn that into a proper test in Piglit...