[AMDGPU] Use v_mad_u64_u32 for IMAD32

Nic Curtis done the experiments to prove it is faster than a
separate mul and add.

Fixes: SWDEV-332806

Differential Revision: https://reviews.llvm.org/D127253