optim: introduce archetypal analysis method & associated test suite

[fel-thomas]

Introduce Archetypal Analysis framework to Overcomplete.

We use the following notation:

• A: data matrix, tensor of shape (n_samples, n_features)
• Z: codes matrix, tensor of shape (n_samples, nb_concepts)
• W: coefficient matrix, tensor of shape (nb_concepts, n_samples)
• D: dictionary matrix, computed as D = W @ A

The objective is:
min_{Z,W} ||A - Z D||_F^2
subject to Z in Δ^nb_concepts and D in conv(A)

Say it otherwise, Z row stochastic and W row stochastic.
Currently supports projected gradient descent (PGD) solver.