First implementation of LogisticRegression model

[NathanRxl]

Part of #4, this implementation of logistic regression supports:

• Binary classification
• Multinomial classification
• Custom penalty

Next immediate improvements in priority order:

• Kernelize LogisticRegression to be able to give a custom kernel (with linear kernel by default)
• Fit intercept
• Parallelize the computations to increase the performance

Co-Authored by: hugues hugues.tavenard@eleves.enpc.fr

Replace heavy computations based on for loops in logloss and its gradient by light computations based on numpy matrix products.

This change increases the running time of fit method by a factor of 100.