[Data] Re-implement APIs like select_columns with PyArrow batch format

[bveeramani]

select_columns, drop_columns, and add_column are implemented as a map_batches with a UDF that uses the pandas batch format.

ray/python/ray/data/dataset.py

Lines 865 to 875 in d5fa9a0

	def select_columns(batch):
	return BlockAccessor.for_block(batch).select(columns=cols)
	return self.map_batches(
	select_columns,
	batch_format="pandas",
	zero_copy_batch=True,
	compute=compute,
	concurrency=concurrency,
	**ray_remote_args,
	)

This implementation has the consequence of converting Arrow blocks to pandas blocks. Because pandas blocks are more issue-prone, we should re-implement these methods with the "pyarrow" batch format.

(Historical context: we needed to use the pandas format because the "pyarrow" batch format didn't work with arbitrary Python objects. With #45272, we no longer have this restriction)