Add refinement to cellular automata to make them actual usable mazes

[sraaphorst]

This deals with issue #23.

After the cellular automata algorithms have run, seldom do they generate usable mazes.

What we should do is determine connected components (using disjoint sets?), and then carve passages between them so that they can actually be mazes. This won't make them perfect spanning trees, but it will help.

[sraaphorst]

Rule B2/S123 seems to do exactly this.

[sraaphorst]

We could still add the refinement, though, which would be useful to ensure connectivity.