Refactor stochastic optimizers

[accosmin]

Simplify & clarify the implementation of the stochastic optimizers:

• have the parameter smoothing optional (see loop.h)
• explicit stochastic & batch problems passed
• always update the batch values (value, gradient) after each epoch
• thus can check for convergence after each epoch using a given epsilon
• extend the learning rate schedule to: fixed, exponential decay, t-inverse (see the SVRG paper - experimentation section)

[accosmin]

Extend test_affine to support other criteria and loss functions. Also check if it is possible to merge accumulator_t and criterion_t. SG/SGM should never fail during benchmark_stoch for convex functions: need a better learning rate schedule?!