nir, ac, aco: Extend load/store_buffer_amd intrinsics with access flags and index source.

This is a pre-requisite for NIR based VS input lowering. These patches can be already reviewed on their own.

• Previously all uses of load_buffer_amd / store_buffer_amd were implicitly treated as coherent. Now add actual ACCESS flags so that their coherency can be configured per instruction. This is needed because VS input loads won't need to be coherent.
• These NIR instructions are compiled to HW MUBUF instructions which have an additional index field which was previously not used. Let's now wire up this field to a corresponding NIR source, because this will be used for VS inputs.
• Additionally some minor cleanup in ACO along with enabling an optimization which was missing for the case when both the offset and index fields are used.