new metrics and losses

[thib-s]

new metrics for robustness estimation need to be added.

• proven robustness radius ( average on the dataset )

also most of the common keras losses have a parameter that need to be tuned (example: temperature in categorical_crossentropy), as these parameters default to 1 in keras there is a need to create a wrapper for such losses.

[thib-s]

wont be for 1.2, rather 1.2.1