[MLton] speeding up type checking by caching

[Stephen Weeks]

The more I think about it, the more I like the idea of speeding up
type checking by transparently caching the result of preprocessing
part of the program.  The semantics is simple -- to the user, it
should be as if the whole program is type checked every time they call
MLton.  But internally, we are free to save worlds that result from
type checking part of the program, and to restart with those worlds if
we can prove that it is equivalent to type checking the whole program.