Various Lengths, Constant Speed: Efficient Language Modeling with Lightning Attention