BEVBert: Multimodal Map Pre-training for Language-guided Navigation