The GPU prefill path can be significantly faster than single-token decode due to fundamental differences in how these two phases utilize