Add argparse parametrization for the finetuning script

[gsarti]

Similar to what is currently available in download_model.py, add Argparse with parameters in finetune_nli.py for parameters:

• model_name, default 'models/scibert', type str

• batch_size, default 64, type int

• model_save_path, default 'models/scibert_nli', type str

• num_epochs, default 2, type int

• warmup_steps, default None, not required

• do_mean_pooling with action='store_true'

• do_cls_pooling with action='store_true'

• do_max_pooling with action='store_true'

Then:

• Add check for only one of the pooling condition to be verified (raise an AttributeError if more than one is). If none is specified, we use mean pooling strategy.

• Check if the warmup_step parameter is set before setting it to 10% of training: if it is, keep the user-defined value.