These bitcasts are no-ops, and lowering prevents some optimizations like 16bit vectorization. So, don't lower_pack anymore for RADV.
v2: added patches to remove any *_split versions from different NIR passes. v3: dropped the NIR patches. (see #3503) v4: added options to lower pack_* instructions via nir_lower_alu_to_scalar(). Swizzled pack instructions cannot be considered bitcasts and are difficult to deal with in the backend.