Added subquestions column with formatted answers

Added diverse 2K samples using image feature clustering

Combined all subsets from textvqa-temp-chunks into single split

initial commit