High-dimensional Penalty Selection via Minimum Description Length Principle