[RFA] Fix glr.c's YY_ERROR_VERBOSE vs. glr.cc's

[Paolo Bonzini]

Currently, glr.cc testing is limited in calc.at because of a couple of 
bugs in glr.cc itself.  These bugs are due to glr.cc's nature of thin 
wrapper around glr.c.

First, glr.cc sets YY_ERROR_VERBOSE to 0 and glr.c interprets it as a 
request to *have* verbose errors, thanks to these directives:

#ifdef YYERROR_VERBOSE
# undef YYERROR_VERBOSE
# define YYERROR_VERBOSE 1
#else
# ...
#endif

This can be considered a bug in glr.c actually, and the same bug is also 
there in yacc.c (if you want, the same glr.c hunk applies to yacc.c as 
well).  This patch fixes it using some preprocessor magic.

Second, yy::parser::set_debug_level is never defined unless %debug is 
specified.  This because it just sets "::yydebug" to its argument, and 
the variable is not defined in glr.c unless YYDEBUG is defined.  This 
patch also fixes it by making the function do nothing unless %debug is 
specified.

Unluckily, the testsuite changes enabled by this patch are tangled in my 
%language work, and I'd rather wait for the copyright assignment.  In 
the meanwhile, these are bug fixes (albeit glr.cc is undocumented) that 
could be applied as uncopyrightable.

Tested with no regressions.

Paolo