Tests for self attention using CumConcatLayer

[Zettelkasten]

Some tests for #391.
I didn't run any of this yet (so most likely buggy) because there is no implementation.
It still needs CumConcatLayer from #589, but also the common arguments for DotLayer from #569.
Also, some way to set dim tags (e.g. set_dim_tags: Dict[str, DimensionTag] for ReinterpretDataLayer. I'm not sure, didn't we add this already somewhere?)

We might also want more tests than this: e.g. some with positional encoding.

[Zettelkasten]

This should be ready now.

I left in test_self_attention_optimize_out, which essentially does the same as test_reclayer_optimize_out_cum_concat_gen_self_att. Should we remove that?

[albertz]

There is also test_generalized_non_rec_self_attention already now.

I did not check in detail, is this really exactly the same, or just similar? What is the difference?
If there is some difference which makes sense for another test, then leave it. Otherwise remove it.

[Zettelkasten]

There is also test_generalized_non_rec_self_attention already now.

Oh okay, I missed that.

The test case in this PR also checks that SelfAttentionLayer does the same as the generalized self attention for the recurrent case (not just non-rec). Maybe we want to check that too.
But maybe we can just close this PR, it's a lot new code which originally was meant to be much more general. I think the test cases we already have are nicer because they use an explicit net dict and seem easier to debug if something breaks in the future.

[albertz]

Hm, I don't know. It might be nice as an example. I would say leave it.