[C++] Vector kernels should implement or reject null_handling = INTERSECTION

As discovered in ARROW-13530, right now the framework will let you register a vector kernel with null_handling = INTERSECTION, but doesn't actually implement that (it'll preallocate but won't compute the result). We should either implement it, or decide it makes no sense and explicitly reject registering kernels with this null handling mode.

**Reporter**: [David Li](https://issues.apache.org/jira/browse/ARROW-16145) / @lidavidm

_{**Note**: *This issue was originally created as [ARROW-16145](https://issues.apache.org/jira/browse/ARROW-16145). Please see the [migration documentation](https://github.com/apache/arrow/issues/14542) for further details.*}

Start by tracing vector kernel registration and the handling of null_handling = INTERSECTION, using ARROW-13530 as the reported context. Determine whether intersection semantics are appropriate for vector kernels; done means either the behavior is implemented correctly or registration explicitly rejects this mode.

Written by the indexing model from the issue text.