Fix a parser bug where we let attributes interfere with our disambiguation

of whether a '(' was a grouping paren or the start of a function declarator.
This is PR2796.

Now we eat the attribute before deciding whether the paren is grouping or
not, then apply it to the resultant decl or to the first argument as needed.

One somewhat surprising aspect of this is that attributes interact with
implicit int in cases like this:

void a(x, y) // k&r style function
void b(__attribute__(()) x, y); // function with two implicit int arguments
void c(x, __attribute__(()) y); // error, can't have attr in identifier list.

Fun stuff.

llvm-svn: 57790