Decompose globals does not flatten globalfunctions when supported

[Dimosts]

As revealed by #370 and discussed in #366, the current version of decompose_in_tree, e.g. [(count([IV1,IV2,IV3],1)) - 2 == 0] stays like that. This creates a problem in solvers that support count only in a direct comparison and that do not use any other transformations later (like minizinc). If flatten is used, like in ortools, this is not a problem, but we should not always have to use flatten to support them.

In my opinion, we should directly flatten these cases in decompose_in_tree. It is easy to do so, using normalized_numexpr(), and offers many benefits.

I will open a PR with this and we can decide if we move towards this or not