Make cases simpler, faster, and easier to understand.

Cases (struct ccase) is reference-counted to avoid copying more
data than necessary.  But the existing implementation that uses
separate structures for references to cases (struct ccase) and
the cases themselves (struct case_data) has a number of issues:

   - The semantics of struct ccase are tricky to remember; one
     has to remember to treat struct ccase as essentially a
     pointer type.  I get confused sometimes myself.

     (One of the symptoms of this was the existence of
      the case_nullify and case_is_null functions, which are
      exactly analogous to setting a pointer to null and comparing
      one against null, respectively.)

   - Every reference to data in a case involves two levels of
     indirection, which is one more than actually necessary.

   - The implementation is somewhat complicated.

This commit simplifies things, by dropping a level of indirection.
Now, a case is just a pointer to a struct ccase, which contains
the actual data.  Reference counting is still present, but it is
simplified.  The only thing that must be remembered is that, before
modifying a case that may be shared, it must be passed through
case_unshare() to make a new unshared copy if necessary.