Don't undo -- just swap with previous

[bchurchill]

Performance improvement.

[jrkoenig]

So I did a little perf test. Turns out there's about a 10% performance penalty for not using undo. It turns out Cfg's are much heavier weight than I thought, with dozens of std::vectors to copy and such resulting in a lot of allocations (which are relatively expensive). If we're not using the incremental recompute code, it might make sense to remove some of this cached data and/or change the layout to be more copy friendly.

This was on examples/{tutorial,parity}, which don't have many test cases and a single basic block so they run quickly (150k-250k iterations/sec) and recompute is cheap. There likely would be less relative penalty for examples like nacl that run at a few thousand iterations/sec and have control flow.