Stream-level composability for state artifacts

As discussed in: https://gitlab.com/meltano/meltano/-/issues/2903

Summary: We'd like to be able to run multiple copies of a tap in parallel and then merge the state from the two jobs. While this likely works in practice with most taps today, it requires some assumptions about how state is stored and where/whether merge conflicts are expected.

For example:

An orchestrator kicks off users and customers sync as two separate jobs, running in parallel. If the STATE output from one job contains a top-level users bookmark, and the STATE output from another job contains a top-level customers bookmark, we would like the two job outputs to be able to be merged together to create a single merged STATE output with both users and customers bookmarks combined.