cluster predicate - unidirectional graph reachability matrix

[LebedevRI]

Questions/notes:
1. Is there any documentation whatsoever on how to add new tests? I'd like to add them, but looks like pytest doesn't test in-tree std, but the install dir
2. Is there an autoformatter?
3. I've only added enum versions. Should there only be int versions, and enum versions shall just dispatch tothose?
4. I'm not sure about the multiroot suffix, should it be anyroot?
5. Also, i'm not sure if cluster is the best name? Maybe constellations / strongly connected components (scc) may be more descriptive, i'm not sure.

I needed this some time back, and having this
would have saved me quite some time.

Essentially, this ended up consisting of two parts:

1. reachable_multiroot() - a seemingly rather boringly straight-forward generalization of the reachable(), which, instead of enforcing that the subgraph is reachable from the root node (given by index, non-optional), and thus forcing there to be a single joint subgraph, instead takes (zero or more) roots as a mask. The rest of the logic is generalized accordingly.

2. cluster() predicate. This one took quite a bit of rewrites to fully distill it to it's purest+simplest+fastest form. :) The (final) basic idea is to, basically, give each node a unique "color", and then set the nodes, linked together by selected edges, to the same color, thus ensuring that the nodes, that are not in any way connected, don't have the same "color", and vice versa.

   Implementation-wise, instead of color, for each node, we track the index of
   the disjoint subgraph to which this node belongs, and it turned out
   that tracking that by tracking the root node of each subgraph is
   the best approach (referred to as root). So there's 3 parts:

   1. Splitting disjoint subgraphs: we then defer to reachable_multiroot() to enfore that each disjoint subgraph actually has a root node, thus different disjoint subgraphs have different roots.
   2. Coalescing same-color subgraphs: nodes, that are linked together by selected edge, have the same root.
   3. And then we actually compute reachability matrix based on whether or not two nodes have same root (belong to the same disjoint subgraph)

   Now, this is controversial. I've decided to return the reachability
   matrix, NOT the subgraph index. While sparse data structure seems better,
   it immediately requires symmetry breaking constraint, which is tricky
   (i have written it.), and --all-solutions actually ends up being MUCH
   slower (O^3?) than just returning reachability matrix...
   The directed version would need to return the matrix,
   so this is also future-proofing for consistency sake

I have an exhaustive test harness for the cluster() (but not the rest), it would not have been possible to refine the predicate so much otherwise, please see https://github.com/LebedevRI/minizinc-graph-reachability-matrix-predicate.mzn/tree/proof-of-concept/unidirectional

Refs. #837

[LebedevRI]

Ok, i think this is basically it. If something else is needed, please let me know.
If this won't go anywhere, at least this was fun.