Add zero-copy conversion from binary to large binary and vice versa

When we write data we will use small/large offsets based on the offsets we receive. In 2.1, and perhaps more in the future, these differences will vanish and we will just write "offsets" and let compression make them whatever size is appropriate.

However, users still should be able to specify what kind of offsets they want to read the data into in memory and this shouldn't be affected by what we stored the data with.

In other words, if we write 4 byte offsets and want to read into 8 byte offsets that should be fine. If we write 8 byte offsets and want to read into 4 byte offsets that should be fine (assuming there is no single value larger than 2GB).

We should allow users to change the data type of the dataset column between binary and large binary without requiring a rewrite of the data.

This also applies for the string types.