Support sliced list arrays in cast

[lhoestq]

There is this issue in pyarrow:

import pyarrow as pa

arr = pa.array([[i * 10] for i in range(4)])
arr.cast(pa.list_(pa.int32()))  # works

arr = arr.slice(1)
arr.cast(pa.list_(pa.int32()))  # fails
# ArrowNotImplementedError("Casting sliced lists (non-zero offset) not yet implemented")

However in Dataset.cast we slice tables to cast their types (it's memory intensive), so we have the same issue.
Because of this it is currently not possible to cast a Dataset with a Sequence feature type (unless the table is small enough to not be sliced).

In this PR I fixed this by resetting the offset of pyarrow.ListArray arrays to zero in the table before casting.
I used pyarrow.compute.subtract function to update the offsets of the ListArray.

cc @abhi1thakur @SBrandeis