ProSST: Protein Language Modeling with Quantized Structure and Disentangled Attention