Artificially augment the training data

[accosmin]

Augment training samples (to reduce overfitting):

• new parameter to all tasks: "noisy=off/on"
• salt & pepper noise (vary the percentage)
• translation, scale, various deformations

info_task should also be updated to display the noisy training samples.

[accosmin]

Extend specific task_t implementations with new parameters to specify if to use noisy inputs and the noise type: noise_type = {off, salt&pepper, deformation}, noise_level = {0, 10}.

This should be available only for the training dataset. The hashing can be made more efficient while at this step: compute it once per sample at loading.

[accosmin]

Create a new program to benchmark the task_iterator_t. Ideally it shouldn't have any significant impact when training models.

[accosmin]

Can also try to create random background samples for classification: random 3D input tensor should be mapped to target {-1}^C, where C = number of classes.