Enforce dim tags to be unique in tensor

[albertz]

The order of axes should never matter.

But when a single dim tag can occur multiple times in a tensor (Data), it does matter. E.g. for operations like SoftmaxOverSpatialLayer on some input [B,T,T].

You can get such tensor e.g. via DotLayer.

We should disallow this, so that the order of axes will not matter.

The user explicitly would need to change one dim tag before to some new dim tag (e.g. via #633 or #589).

This would be helpful for #391.

[albertz]

This obviously introduces new behavior. So it could be introduced via a new behavior version (#508).

However, the question is whether this behavior should be enforced on everyone. A new behavior version usually implies for an easy transition of old code to new code. But this is a case where it might be a bit non-trivial in some cases.

Instead, we could introduce this as a behavior flag. Some config option like behavior_unique_dim_tags = True or so.

[albertz]

We now stumbled upon a case where there is no clear obvious solution on how to avoid ambiguous tags in a tensor, namely for the weight matrix of linear transformations with in_dim == out_dim. In this case, it would be a (dim, dim) weight matrix.

This is also shown here: https://github.com/rwth-i6/returnn_common/issues/17#issuecomment-997312167

One potential solution was to introduce some match_priority: #871