Faster map w/ input_columns & faster slicing w/ Iterable keys

[norabelrose]

@lhoestq Fixes #2193

• map now uses with_format to only load needed columns in memory when input_columns is set
• Slicing datasets with Iterables of indices now uses a new Table.fast_gather method, implemented with np.searchsorted, to find the appropriate batch indices all at once. pa.concat_tables is no longer used for this; we just call pa.Table.from_batches with a list of all the batch slices.

Together these changes have sped up batched map() calls over subsets of columns quite considerably in my initial testing.

[lhoestq]

Thanks a lot !
This looks all good to me :)

Could you just run make style to format the code and make the CI green ? :)

[norabelrose]

@lhoestq Just fixed the code style issues— I think it should be good to merge now :)

[lhoestq]

Thanks ! LGTM

We may need to improve the case when input_columns is used on a dataset formatted with output_all_columns=True. Maybe this is something we can take care of in another PR