Column Selection: Improvements

[ritugala]

Description

• Tried to add information about columns from huggingface, but I couldn't find anything super relevant. So I decided to just add the dataset description instead. Made the appropriate changes in the base prompt as well.
• Updated the prompt for column selection by (1) adding an example where there is a mismatch between the task and dataset chosen, so there are no relevant columns (eg using a machine translation dataset for a summarization task), and (2) swapped an old example for another one
• Fixed bug in truncate_row -- only add the "..." if len(row) > max_length
• Added support for datasets like opus100 (which have nested columns), so that these datasets can be used and we don't try to generate a dataset -- the solution was pretty straightforward, using flatten() function of huggingface + some trivial column preprocessing

[zhaochenyang20]

Generally well-structured

[viswavi]

Looks good to me, just left two very minor suggestions! Feel free to merge after making these changes.

[viswavi]

What will happen if there is a column conflict here? E.g. we have two columns, foo.baz and foo_baz? Then I think these will conflict and one of the columns will be erased.

I know this is unlikely but might be good to check and handle this scenario.

[viswavi]

@ritugala LGTM. Please test this in the CLI demo before merging.