Refactor graphchain

[lsorber]

Resolves #16, #19, #30, #31, #32.

Additional features in this PR:

• If two keys map to the same computation, fetching both these keys will result in only one of them actually being computed while the other is loaded from cache (cf. example_5.py).
• The previous point also applies recursively, i.e., two computations are "the same" even if they depend on different keys, as long as those keys result in the same output (cf. example_5.py).
• Caching is no longer done by default, but only if it saves time. The definition of saving time depends on the time it would take to compute the key and load the necessary dependencies vs. the time to load the key from cache.
• Adds a log trail to the cache folder. Can later be used to do housekeeping.

[zgornel]

Looks great; did not have enough to test it extensively however so far so good :). It may be time to merge and leave the old implementation on a legacy branch or something for reference (if it makes sense).