[COG-946] Abstract away input dataset

[alekszievr]

Currently evals/eval_on_hotpot.py handles only the hotpotQA dataset. Make the dataset choice an input of the eval. Prepare for hotpot and https://github.com/Alab-NII/2wikimultihop and subsets of them.

Steps:

• implement loader for hotpotQA-compatible datasets
  • including 2wikimultihop
  • write documentation of accepted file formats
• define a list or dict of acceptable datasets
• add a "dataset" argument to the eval

_{From SyncLinear.com | COG-946}