Allow `DialogFlowScore` to work at sub-turn level

[sergioburdisso]

Allow the user to optionally specify the embeddings won't be for the whole turn but rather for the last utterance (sentence). Internally, each turn.text should be split into sentences (e.g. nltk.tokenize.sent_tokenize()) and embed only the last one (as in Section 2.2 of https://aclanthology.org/2023.findings-emnlp.492.pdf).