Improve minibatch optimizers

[accosmin]

Do not create the optimizers for each minibatch round to save memory and improve speed.

Possible solution:

• fn_fval: use the current minibatch
• fn_grad: move to the next minibatch if (iteration % epoch_size == 0)

Consider also tuning the number of iterations per minibatch {1, 2, 3, 4}...

[accosmin]

Need to reuse the optimization context (if possible):

• GD: need to clear/reinitialize the linesearch initialization algo after each minibatch
• LBFGS: clear the history correctly when the line-search algorithm fails
• CGD: clear the previous state when the line-search algorithm fails