add last-layer finetuning for tasks

[yongzx]

• add arguments parser.add_argument("--task_layers", choices=task_layers, required=True).
• introduce task-specific finetuning only on the last layer of the Transformer model.
• Resolved #24.

[yongzx]

Further commits:

33 - remove a line of assert False.

e1079c1 - use correct Trainer class, and use model.freeze_model(freeze=True) to freeze the base model instead of using for-loop to freeze base model. The reason is that the latter would break if we use BERT model that has "bert" prefix instead of "transformer" prefix.