drm/amdgpu: use scheduler user seq instead of previous user seq