scGPT: use sparse matrix calculations

[dorien-er]

see #754 (comment)

This will expand the sparse matrices into dense matrices again, which unfortunatly will cause out of memory issues. Especially when working at atlas scale.

Looking at the implementations for tokenize_batch https://github.com/bowang-lab/scGPT/blob/706526a76d547de4ed711fa028c99be5bdf6ad8a/scgpt/tokenizer/gene_tokenizer.py#L249

and pad_batch https://github.com/bowang-lab/scGPT/blob/706526a76d547de4ed711fa028c99be5bdf6ad8a/scgpt/tokenizer/gene_tokenizer.py#L312

I feel that an implementation that does not expand the matrices into memory is possible to implement. To be discussed.