Add cache for more efficient refs

[rdvdijk]

Using ref is sometimes extremely expensive when referencing values parsed very "early" in large ParseGraphs. This is a common pitfall in many tokens. We can improve this by creating a cache for all parsed values, which is used by evaluating a ref.

An example: the CFB format contains a 'sector size' in its header, which is used throughout the token. The lookup of this sector size becomes extremely slow in CFB with large sector chains.