glr2.cc: start the transition to using symbol_type

Currently glr2.cc uses three variables/struct members to denote the
symbols' kind (or state), value and location.  lalr1.cc has two types
for "complete" symbols: symbol_type and stack_symbol_type.  Let's use
that model in glr2.cc too.

For a start use yyla (a symbol_type) to denote the lookahead, instead
of the triple yytoken, yylval and yylloc.  This will make easier the
introduction of the "context" subclass, used in parse.error=custom.

It simplifies the code in several places.  For instance from:

    symbol_kind_type yytoken_current = this->yytoken;]b4_variant_if([[
    value_type yylval_current;
    ]b4_symbol_variant([this->yytoken],
                       [yylval_current], [move], [this->yylval])], [[
    value_type yylval_current = this->yylval;]])[]b4_locations_if([
    location_type yylloc_current = this->yylloc;])[

to:

    symbol_type yyla_current = std::move (this->yyla);

* data/skeletons/glr2.cc (yytoken, yylval, yylloc): Replace by...
(yyla): this.
Adjust all dependencies.
(yyloc_default): Remove, unused.
* tests/c++.at, tests/glr-regression.at, tests/types.at: C++11 is
required for glr2.cc.
Adjust to changes in glr2.cc.

tests/c++.at

@@ -1031,6 +1031,7 @@ yylex (yy::parser::value_type *lval)
 AT_BISON_CHECK([[-o input.cc input.yy]])
 
 AT_FOR_EACH_CXX([
+  AT_GLR2_CC_IF([AT_REQUIRE_CXX_STD(11, [echo "$at_std not supported"; continue])])
   AT_LANG_COMPILE([[input]], [[input.cc scan.cc]])
 
   # Leave enough valid tokens to make sure we recovered from the