SLAB: Efficient Transformers with Simplified Linear Attention and Progressive Re-parameterized Batch Normalization