Data: sparse_dim: DimensionTag to unify sparse, dim, vocab

[albertz]

It would be nice to have a DimensionTag which describes the sparse dim (when Data.sparse is True).

[albertz]

DimensionTag currently does not contain vocab but we can move it there.

[Zettelkasten]

This is also nice/required for when we add an axis/in_dim argument to LinearLayer. (#597)

[Zettelkasten]

Would this be part of Data.dim_tags, or not (because there can only be one of these tags for each Data instance)?
Maybe it's nice if a sparse dim is just another dim like all others. But I guess many layers would need special handling for this dim tag, so maybe it makes sense to make this more explicit. I'm not sure about the implications.

[albertz]

Data.dim_tags corresponds to the dims which are really present in the tensor. So to batch_shape. Or we have the property Data.placeholder.ndim == len(Data.dim_tags). I would leave that behavior of dim_tags.

However, we can make a separate implicit_dim_tags or so. Very similar as what I proposed as out_implicit_shape in #706.