Check model lipschitzness

[franckma31]

Add several empirical evaluations of the Lipschitz constant of a model.
Code is also moved from utils/init.py to utils/evaluate_lip_const.py

Methods implemented are:

• jacobian_norm: evaluate the norm of the jacobian ($||\nabla_x f(x)||$) on a set of input samples
• noise_norm: evaluate random noise influence on the output ($||f(x+\delta)-f(x)||/||\delta||$)
• adversarial attack: Same evaluation, but performed on the attack that maximizes the Lipschitz constant $||f(x+\delta)-f(x)||/||\delta||$ using the Projected Gradient Descent (PGD) algorithm. Additionally, a two-step attack is introduced to compute $||f(x+\delta_1+\delta_2)-f(x+\delta_1)||/||\delta_2||$ allowing us to search for a steeper slope in the neighborhood of $x+\delta_1$

[cofri]

This is very useful!
I left some comments