Improve stochastic optimizers

[accosmin]

Need to improve tuning their hyper-parameters because there is a BIG variance in the resulting train/test error.

Possible solutions:

• limit the decay rate: [1e-4, 1e-1] to make sure some progress is performed
• tune on larger number of epochs to have consistent results

[accosmin]

There might be a bug in tuning the stochastic optimizers: the selected configuration does not correspond to the minimum training loss.

[accosmin]

Fixed the bug related to tuning. Many more samples are used now for training, thus hyper-parameter tuning is much more robust.

[accosmin]

Fixed all stochastic optimizers except AG-based ones.

[accosmin]

Fixed AG-based optimizers as well.