Use interior mutability to represent graph predecessors

[Wilfred]

See the mutable_vertices branch. This might save perf, as we're currently spending most of our time looking up items in hashmaps.

It's also worth exploring whether a conventional eager graph construction is better than the current lazy model.

Once that's done, we should look at running each diff in parallel rather than sequentially.

[Wilfred]

Could pre allocate the simple nodes ignoring parents, or try to find a later node that came after, or grouping seen by syntax ID pairs.