backwards support for pre-py3.10, add datasets requirement used in train

[winglian]

match is python 3.10+, this should let others use this in environments like colab datasets is required for this import: https://github.com/johnsmith0031/alpaca_lora_4bit/blob/main/train_data.py#L3 sentencepiece is required for LlamaTokenizer: ImportError: LlamaTokenizer requires the SentencePiece library

[johnsmith0031]

Thanks you for the fix!