Decouple vocabulary from BMX, BM25 models

[juliuslipp]

Problem

• Sparse index models (BMX, BM25) maintain the vocabulary to map tokens to token ids, while building sparse matrices only needs token ids.
• BaseSparseIndex stores the corpus tokens for index rebuilding. It takes up a lot of memory if the corpus is big.

Suggestion

• Maintain the vocabulary and perform token mapping in BaseSparseIndex to simplify the model logic.
• Save token ids for corpus instead of tokens for index rebuilding to save up memory, as only token ids are necessary for matrix construction.