glr.cc: fix the handling of syntax_error from the scanner

Commit 90a8537e62 was right, but issued
two error messages.  Commit 80ef7e7639
tried to address that by mapping yychar and yytoken to empty, but that
completely breaks the invariants of glr.c.  In particular, yygetToken
can be called repeatedly and is expected to return the latest result,
unless yytoken is YYEMPTY.  Since the previous attempt was "recording"
that the token was coming from an exception by setting it to YYEMPTY,
instead of getting again the faulty token, we fetched another one.

Rather, revert to the first approach: map yytoken to "invalid token",
but record in yychar the fact that we come from an exception thrown in
the scanner.

* data/skeletons/glr.c (YYFAULTYTOK): New.
(yygetToken): Use it to record syntax errors from the scanner.
* tests/c++.at (Syntax error as exception): In addition to checking
syntax_error with error recovery, make sure it also behaves as
expected without.

data/skeletons/glr.cc

@@ -284,6 +284,7 @@ b4_percent_code_get([[requires]])[
     /// \returns  0 iff parsing succeeded.
     virtual int parse ();
 
+#if ]b4_api_PREFIX[DEBUG
     /// The current debugging stream.
     std::ostream& debug_stream () const;
     /// Set the current debugging stream.
@@ -295,8 +296,8 @@ b4_percent_code_get([[requires]])[
     debug_level_type debug_level () const;
     /// Set the current debugging level.
     void set_debug_level (debug_level_type l);
+#endif
 
-  public:
     /// Report a syntax error.]b4_locations_if([[
     /// \param loc    where the syntax error is found.]])[
     /// \param msg    a description of the syntax error.