Performance improvements

Currently it is hard working on very large programs, as there is some slowdown on connect/disconnect. This is due to:
1. Redrawing blocks is slow.
2. Algorithm W is exponential in time (empirically, I think). This is further worsened by Type.apply applying every substitution, which is required for fully substituting over some of the 'incomplete' types tagged.

Some ideas:
1. Only redraw blocks which had a change of type. The current behavior is to re-render the connected/disconnected blocks and their parents.
2. When composing substitutions, fully apply them. I'm not sure if this will mess up intermediate substitutions.
3. Maybe cache things that don't change too often.

(I'm opening a new issue instead of reopening #177 as the implementation details have changed)