LRU cache

Thunks can now be slated to be cached using the cache kwarg. Once a cached thunk is evaluated, the result will be kept around in the worker's memory until Dagger.MAX_MEMORY[] is exceeded by all tracked objects in memory, at which point the results are dropped from cache in an LRU order. Thunks whose results get dropped will be re-evaluated. This feature is not used by Dagger's array library yet.

This release also has some minor optimizations to scheduler bootstrap code.