Implement better control over staged-out datasets

Currently, every dataset in the namespace is staged out (or served in xcube, if the container is run in server mode). The simplest workaround is to del any datasets that should be ignored, but this can get awkward if there are too many of them. xcengine should offer an alternative output selection to let the user explicitly choose which datasets are to be exported. A collection with a reserved name (xce_output_datasets or similar) would be easy, but would require reading documentation to use. Maybe there is some more discoverable alternative.

Start by examining how xcengine currently stages every namespace dataset and how server mode serves them. Decide on a discoverable way for users to select output datasets, then verify that non-selected datasets are neither staged out nor served. The issue names no files or tests, so the relevant entry points must be located first.