tolerance: absolute or relative.

[rainwoodman]

The tol argument of LBFGS is relative, but GradientDescent is absolute. Does it make more sense to use absolute tolerance -- for we have been using convergence condition based on drop of chi square < 1, not relative drop of chi square.

@aslanyan