Allow inserting subschemas

[wjones127]

If a dataset has columns a, b, c, a user should be able to call dataset.write(..., mode="append") with columns a, b, and it will fill c with null.

Physically, we should allow the new fragments to not include c at all, to save space. This would also open the door for inserting super schemas (which would expand the schema).

Potential follow ups:

• Support inserting in merge_insert with a subset of columns #2904
• Support add columns with all nulls #3044