support training bloom-560m and load dataset from local dir

[hzg0601]

1. support load dataset from local path in offline mode
   For those who can't access huggingface hub stably, they have to download the repo of huggingface hub manually and
   upload dataset in the offline mode. However, current implementation use the model_name as the key to load corresponding dataset class of raw_datasets.py which would raise RuntimeError( f"We do not have configs for dataset {dataset_name}, but you can add it by yourself in raw_datasets.py." ). This PR fix the problem by split the string of local path.
2. Causally, some repo owners of huggingface hub, such as bigscience, would like to separate model and tokenizer. So we have to add a tokenizer_name_or_path option in args in case of that situation.

[hzg0601]

@microsoft-github-policy-service agree

[conglongli]

@hzg0601 Before reviewing this PR, we need your help on two things:

(1) Why your PR is deleting this log file? https://github.com/microsoft/DeepSpeedExamples/blob/04b1036349b0d02e0f940c8431414ef18e96ddf2/applications/DeepSpeed-Chat/training/step1_supervised_finetuning/training_log_output/opt-1.3b-globalBatchSize128.log

If this is by mistake, please restore it. If this is not by mistake, please give a clear reason of why deleting it.

(2) The formatting test is failing. Please follow my comment at here #597 (comment) to fix it.