On the token distance modeling ability of higher RoPE attention dimension