[AMDGPU] Use VGPR classes in divergent build_vector patterns

This does not appear to affect codegen, but using SGPR classes for
operands of VALU instructions looked weird.

Differential Revision: https://reviews.llvm.org/D136459