Implementing #120

Deprecating ModelList in favor of ModelMapping.
Deprecating the use of a Dict for model mapping in favor of ModelMapping.

ModelMapping is now the unique interface users face for declaring models.
At the time the single-scale pathway creates a ModelList, so we re-use all previous code. The multiscale pathway re-uses the previous Dict one.

This PR is not about reducing the pathways to one, but future work should do that, i.e. one way for computing the dependency graph, making all checks, etc, whatever the scale (single- or multi-). The only major difference I think is that ModelList can pre-allocate its outputs because their length is fixed. This would require some rework in the multirate setup, though (see below).

For now multirate is not handled in ModelList, but we should make a multiscale setup from the ModelList in this case (as a shortcut) before implementing something more clever that may pre-allocate outputs.

To do:

• Make the tests pass (use ModelMapping for user-facing code (e.g. doc and tests), and ModelMapping{SingleScale} when possible in the source code, or ModelList otherwise.
• Check that multirate is well handled in ModelMapping{SingleScale}: should either pre-allocate properly (involving) or at least fall back to the multiscale approach. In the latter case, handle the case where we don't provide the MTG (no MTG in single-scale setup)