Faster tunning of the stochastic solvers

[accosmin]

• rationale: can use fewer updates to decide which hyper-parameters are good enough
• solution (1): modify stoch_params_t to specify the number of tuning iterations (default = epoch_size)
• solution (2): modify stoch_trainer_t to use epoch_size / 10 as the number of tuning iterators
• solution (3): may modify stoch_trainer_t to also evaluate a configuration using fewer samples (e.g. training samples / 10) - to check if feasible