New rx implementation with extension constructs

[Mattias Engdegård]

The rx regexp notation is nice to use but the implementation isn't wonderful; 
there is a proposed replacement rewritten from the ground up. It is cleaner, 
has fewer bugs, and is maybe twice as fast.

Most importantly, there is now a proper extension mechanism: for global 
definitions,

 (rx-define snobol-identifier (seq alpha (0+ alnum))

which are available anywhere, and local ones,

 (rx-let ((natnum (1+ digit))
          (integer (seq (opt "-") natnum)))
   ...body...)

where a set of definitions are only available in a lexical scope. This 
zero-cost construct can be placed inside a function, or at top-level enclosing 
multiple variable and function definitions, all sharing the same named rx forms.

Both rx-define and rx-let admit two kinds of definitions:

 NAME RX-FORM
 NAME (ARGS...) RX-FORM

for plain rx symbols and for parametrised forms, respectively. For example:

 (rx-let ((name (1+ letter))
          (comma-separated (x) (seq x (0+ "," x))))
  (rx (comma-separated name)))

works just as expected. &rest arguments are permitted, and expand to implicit 
(seq ...) forms.
No provision was made for macros able to execute arbitrary Lisp code; I just 
couldn't find a use for them, and decided to wait until someone would tell me 
otherwise. Thus, all parametrised forms work by plain substitution.

The code currently resides at https://gitlab.com/mattiase/ry; it will naturally 
be renamed to `rx' once it's in the Emacs tree. It can be integrated in a 
separate branch of the Emacs source repo if you wish, or as patches if you 
prefer that for reviewing. The diffs don't make much sense since it is a 
reimplementation with very little in common with the old code.

The exact form of the extension mechanism isn't set in stone, and I'd welcome 
any suggestions for improvement.