Speed up parsing by storing intermediate tokens

[earwig]

There are certain situations where we rapidly construct and destroy and then have to reconstruct the same token stack when hitting bad routes. These are rare, but act as a major slowdown when they do occur, potentially hitting the parse cycle limit. Ideally, when failing a route, we store the tokenization of complete nodes in some kind of cache which associates them with the starting/ending head locations and the context. This cache can then be popped from if we reach that head location again with the same context, allowing us to bypass regular parsing.

[earwig]

🎉 💯