Extend solvers

[accosmin]

• parameters for stochastic solvers: lambda = auto => means should tune it, otherwise use a finite_space_t{value}
• extend apps/benchmark_[batch|stoch] to filter solvers to benchmark (e.g. --solvers , --solvers ada, --solvers sg,sgm done
• extend apps/train to pass specific parameters for solvers
• for ML allow smaller decay (e.g. 0.0, 0.1 ...)