What anisotropy in embedding spaces is, why all sentences seem 'similar' in raw BERT embeddings, and how it degrades retrieval quality.