Add flag for computing topological depths of neurons before post-processing

[neumannjan]

Added an optional flag (disabled by default) that allows computing topological depths of neurons even without dropout enabled.

Found (and fixed) a bug that when the depth computation is run, ISO value compression then fails.

[GustikS]

Nice! but adding new fields to the base neuron class just for this is an overkill, that's why I was reusing the depth field - there are millions of neurons and we want to save every bit of space (perhaps let's discuss on Thursday). Thanks!

[neumannjan]

In that case a wrapper class could be used for the neurons when running the topological ordering – that way the "topoState" field would exist only throughout the duration of the ordering algorithm itself.

The depth field cannot be reused since it is already used for the actual layer values. By reusing the layer field, you not only overwrite the layer data, but the iso compression breaks during the topological ordering because the (layer == DEFAULT) check returns false for nearly all neurons (even though it is supposed to return true for all of them at the start of the topological sort).

However, it is possible that the layer value computation will not be needed

[GustikS]

Great! Yes, I agree, a wrapper or hashmap or something like that living in the local scope only. Thanks!