STS task improvements

[fdalvi]

For after the deadline, just suggestions and open to discussion:

• Track1 and Track2 datasets can be collapsed into one? If not, maybe we can have a shared base class if there are some differences (e.g. citations?)
• Train data args should be a list instead of manually pre-processing into the *.all file for Fewshot
• Can we add input_ids for de-duplication?

@MaramHasanain @firojalam