Improve line searches

[kellertuer]

currently linesearches like ArmijoLinesearch but also the Wolfe ones and the NonMonotone one are a bit unstable for the simple reason that the involved while loops in the common linesearch_backtrack is maybe a bit too academic and might have an infinite loop.

Ideas

• allow for a minimal stepsize when to stop (already partly available but not yet set as a default=
• allow for a maximal number of steps
  • (Linesearches.jl has 1000 src
  • Manopt/Matlab has 25 src works on a cost evaluation budget)
• check whether we hit NaN/Inf as LineSearches src

And of course one question is then what to do – LineSearches errors, while Manopt just continues after 25 and if the step is too small it is also a question what to do

• error
• continue as is
• have a stopping criterion that stops in that case and reports that problem as a stopping reason (a non-convergence indicating of course)

[kellertuer]

A second thing is to check the defaults, the factor when backtracking, the sufficient decrease factor and such.