Add the quadruple-precision constants; testvect works better now
on sparc64.