Implicit knowledge

[xeren]

Reduces the memory usage of the Relation Analysis. The initial knowledge of base relations like ext, int, po and X*Y is not explicitly stored. This comes with hopefully-negligible runtime costs.

[ThomasHaas]

Performance-wise, this PR seems to give a slight degradation (from 170s to 200s on BBQ, subject to fluctuations). Memory-wise, I haven't checked the difference. It would be good to compare the number of implicitly represented tuples vs. totally represented one's to get a quantitative idea of how much improvement there is.

Apart from the above, I think the RelationAnalysis is becoming far too complicated. The current version is already missing some high-level structure/description, but with all the new implicit representations, it just gets out-of-hand. I think we need a more properly engineered solution that is better maintainable.