No need to build the benchmark suite 4 times.

This is a "dumb" update in that we retain exactly the same functionality
as before.

This adds a couple new methods on HIR expressions for determining whether
they are literals or not. This is useful for determining whether to apply
optimizations such as Aho-Corasick without re-analyzing the syntax.

Finally, if a regex is just `foo|bar|baz|...|quux`, we will now use plain
old Aho-Corasick. The reason why we weren't doing this before is because
Aho-Corasick didn't support proper leftmost-first match semantics. But
since aho-corasick 0.7, it does, so we can now use it as a drop-in
replacement.

This basically fixes a pretty bad performance bug in a really common case,
but it is otherwise really hacked. First of all, this only happens when a
regex is literally `foo|bar|...|baz`. Even something like
`foo|b(a)r|...|baz` will prevent this optimization from happening, which
is a little silly. Second of all, this optimization only kicks in after
we've compiled the full pattern, which adds quite a bit of overhead. Fixing
this isn't trivial, since we may need the compiled program to resolve
capturing groups. The way to do this is probably to specialize compilation
for certain types of expressions. Maybe.

Anyway, we hack this in for now, and punt on further improvements until
we can really re-think how this should all work.