Avoid redundancy in Workflow JSON

[forman]

Expected behavior

When serializing a workflow to JSON, Cate should not generate input entries for operations steps whose input values are either not given or are equal to the operation input's default value.

Actual behavior

Input entries are always generated for every input. For example a step that invokes read_netcdf with just the file input will look like::

            {
                "id": "p",
                "op": "cate.ops.io.read_netcdf",
                "input": {
                    "file": {
                        "value": "this/is/the/path/to/my/file"
                    },
                    "drop_variables": {},
                    "decode_cf": {},
                    "normalize": {},
                    "decode_times": {},
                    "engine": {}
                },
                "output": {
                    "return": {}
                }
            },

This forces incompatibilities if an operation's input name or type changes in the future. For example, I just renamed the harmonize input into normalize for operation open_dataset`. Now all "old" workspaces don't work anymore.

Steps to reproduce the problem

1. Create a workspace
2. Add some step with multiple defaults, e.g. read_netcdf
3. Save workspace myspace
4. Inspect myspace/.cate-workspace/workflow.json

Specifications

Cate 0.9.0dev3

[forman]

When Save Workspace is called from GUI, source inputs now erased in JSON. Workspace is corrupted and cannot be reopened.

[forman]

Should also drop ".return" prefix of source values.