Add sentence-transformers CrossEncoder support

fix(reranker): avoid re-computing the first batch in predict()'s batch-size probe to reduce additional computational effort (#1)

Update kalm_reranker.py

add modeling file