fix: integrate field coords and object arrays for changing field shapes

[eckelsjd]

fixes #31, fixes #32, fixes #33

Proposed Changes

• System and Component predict functions handle the special {{ var }}_coords variables -- which are the field coordinates returned by a model associated with the {{ var }} field quantity.
• Field coordinates and field values are now returned by Component.call_model as numpy object arrays, where each element is the field quantity itself for a given input sample. This allows variable shapes/sizes for field quantities between inputs to be stored in the same data structure.
• Compression now handles interpolation to/from grids where field coordinates and values are passed as object arrays. This assumes 1d object arrays, and will only loop over the first axis when interpolating.
• As a small unrelated fix, you can specify domains as a [ list ] in yaml for scalars, which will get converted to tuples internally as long as the length of the list is exactly equal to 2.
• Another tacked-on fix -- YamlLoader will now fail if it can't find the specified yaml file. It will try to load from a path; if this fails, then it will try to load from a stream directly.