[ENH] Enable output from Scivision model prediction to be used as inputs for additional model

[evangeline-corcoran]

From discussions with users in the MapReader and AI for Biodiversity communities, there is a clear need for Scivision to allow for and facillitate multi-stage image analysis pipelines that might include more than one machine learning model.

For example, a user may first want to identify patches from maps with buildings using the mapreader model, then secondly use a different model to classify the shapes of the buildings.

To allow this the user would need to be able use the output of the first model as inputs in a subsequent model. This might require a change to what's returned by scivision model predict, potentially requiring stricter limits on what is returned to make sure the format is compatible with other models

[ots22]

Related #208