AMDGPU: Don't use separate cache arguments for s_buffer_load node

There's not much value to this separate node from the intrinsic. Make
the operand structure the same as the intrinsic, so we can reuse the
same pattern for GlobalISel.