Make shape and dims available?

[albertz]

Currently, given some LayerRef (which can be thought of as a normal tensor, when comparing our code conceptually to PyTorch or TensorFlow), we cannot really get any information about it (despite the layer ref), such as its dtype or shape.

It would not be too hard to make it available though because all RETURNN layers have get_out_data_from_opts where this can be inferred. And this function get_out_data_from_opts is by design not involving any TF operation, or adding anything to the current active TF computation graph. (There are some smaller technical things to be considered with this approach but they are all easily solvable.)

Having this would allow to more easily check e.g. if the input is sparse (maybe for #38), maybe reusing or checking for specific dim tags, etc.

We can either make just the RETURNN Data instance available as-is (LayerRef.data), or provide some more simple wrappers, like LayerRef.shape = data.dim_tags or (or with #48, more like LayerRef.shape = set(data.dim_tags)) or so.

This is also very related to #17.

[albertz]

We have that now. See e.g. _move_layer_ref_to_new_name_ctx.