Initial support for "transforming" plugins, incl. dbt

[mildbyte]

• Add a TransformingDataSource mixin that lets a data source define what images it requires to run, with an ImageMounter class that can mount and give it links to the schemas these images are mounted in at runtime.
• dbt_utils changes:
  • Let the sources.yml patcher take in a dbt source <> schema map and overwrite sources with required schemas (instead of the current behaviour where we point the source and the target to the same schema)
  • Add a function that runs the dbt project in compile mode and extracts its manifest (https://docs.getdbt.com/reference/artifacts/manifest-json) to list models that the project can build
• Add a DBTDataSource that runs dbt transformations from Git:
  • introspect() gets the repository's manifest and gets a list of all models
    from it that materialize as tables, allowing the user to select models
    that will get loaded into their Splitgraph repo.
  • The plugin requires a map of dbt source names to Splitgraph images (uses
    a JSONSchema with an array)
  • load() uses the new ImageMounter functionality to build the source -> temporary schema map and calls the dbt
    wrapper that injects the Splitgraph image references instead of dbt sources and builds the required tables

Tested against a real-life project at https://github.com/splitgraph/jaffle_shop_archive/tree/sg-integration-test

Not implemented yet:

• Exporting dbt dependencies as provenance (and in general, unifying this with some Splitfile execution mechanisms)
• Graceful error handling, e.g.:
  • scanning the manifest/project for sources that aren't mapped to Splitgraph images
  • scanning for dangling ref() invocations that don't map to any models/sources