Stacker requires vertical levels are [0, ..., n]

[brianhenn]

Enforced here. This prevents training ACE on data where the levels indices start at 1, or are an arbitrary sequence of consecutive integers. For example, taking our current vertically-coarsened atmosphere datasets like ERA5 and training on levels 1-7 (removing stratosphere) is not possible without fully reprocessing the dataset.

This is probably because when establishing the vertical coordinate from the dataset we also assume that eg a_ks and b_ks are zero-indexed, so that the coordinates and integrands align in vertical integrals.

We could relax these requirements and allow increasing sequences of integers for the level names in the stacker, and index-based getters of a_ks and b_k values in the vertical coordinate.