[solver] epsilon Clarke gradient test for quasi newton methods for non-smooth problems

[accosmin]

• quasi-newton methods are surprisingly accurate for non-smooth problems
• but there is no evident stopping criterion
• as such use the Clarke gradient to check for convergence when the function value updates are smal enough.
• see the stopping criterion used in "Nonsmooth optimization via quasi-Newton methods"

[accosmin]

bundle methods do already this