Add possibility for tensors with 'preferred' dimensions

[michi42]

E.g. for long time series, which you know you're not going to slice them in the time dimension. With the current API, this creates huge tensors which are rather slow to process. One could imagine to have e.g. a small tensor (for the other dimensions) of collections/lists (for a 'preferred' dimension, e.g. time).

tensorics-ext-cern CircularBufferTensorbacked uses a similar approach, but has to flatten down to a regular tensor eventually...