Building upon #110, this restores the old behaviour of the result being nullable
when root level errors are present.
While the result is guaranteed to not be nullable in standard conditions (since
either result or errors are always present), when errors are moved to the root
level it could become null, so declaring it non-nullable propagates the null up
to the data field.
This actually causes compatibility problems with some client libraries (e.g.
Relay) that expect the inner result to be null, _not_ data, if there's an error.
This also adds dedicated RootLevelErrors versions of the Api and the Schema
since the configuration is accessed at compile time now, so put_env was not
enough to test them correctly.