Make regexp/no-dupe-disjunctions account for nested alternatives

[RunDevelopment]

This fixes #402.

---

To implement this, I needed to introduce 2 new concepts to the code base:

1. Nested alternatives: These are alternatives or things that behave like alternatives (e.g. character class elements). They are called nested because they are the (transitive) child nodes of some root alternative.
   E.g. if a(c|b)[de] is the root alternative, then b, c, d, and e are nested alternatives.
2. Partial NFAs: The partial NFA of a given root alternative and nested alternative is the NFA that contains all paths of the root alternative that go through the nested alternative.
   E.g. for the root alternative (a|b)(c|d|e)f and nested alternative d, the language of the partial NFA will be (a|b)df.

These 2 concepts are useful because they allow us to ask: Is there some part of an alternative that is a subset?

This is actually quite a bit more general than simple de-sugaring of character classes. This PR allows use to detect partial subsets and duplications within alternatives, and that includes character classes.

---

Side note: I expected this to massively impact performance. We are potentially creating a bunch more NFAs and DFAs than before, but (surprisingly) I couldn't measure any difference when I ran this PR against Prism's 2k regexes. (It did find 5 bugs though.)

[ota-meshi]

Thank you for your wonderful work!!
I made two comments.

[RunDevelopment]

Sorry for the inconvenience. I just found a bug in PartialParser that caused it to construct the partial NFA incorrectly for alternatives. I should have tested this as well.

[RunDevelopment]

Again, sorry for the inconvenience, @ota-meshi.

I think I'm done. Could you please review?

[ota-meshi]

LGTM!