Nuclear norm implementation

[kanxu526]

(1) I fixed a few issues with the nuclear norm implementation, e.g., the original code won't work in the 1-dim case (because the parameter is passed as a 0-dim vector).
(2) I revised the penalty calculation of nuclear norm and pass a new arg called coefsize to reshape the coef matrix and pass to nuclear norm. This is useful for matrix sensing like Negahban & Wainwright 2011. To my knowledge, it hasn't been applied in any of the scikit-learn related toolkits.
(3) Not sure what data the original nuclear norm can be applied to. Does that work similarly as 'multiclass' in sklearn? I failed to find an arg like that in the codes.

I'm quite new in constructing a module and so far I've been working on it just for my own research use. Let me know if you like this idea or not.

[idc9]

Here is an example of using the nuclear norm -- we are working on better documentation. You may want to install the latest version (made some fixes so the below code works!)

import numpy as np
from yaglm.Glm import Glm
from yaglm.GlmTuned import GlmCV
from yaglm.config.penalty import NuclearNorm
from yaglm.config.flavor import NonConvex

n_samples = 100
n_features = 10
n_response = 5

# sample data
np.random.seed(1)
beta_true = np.ones((n_features, n_response))
X = np.random.normal(size=(n_samples, n_features))
noise = np.random.normal(size=(n_samples, n_response))
Y = X @ beta_true +  noise


# some examples
single_pen_val = Glm(penalty=NuclearNorm(pen_val=1.5)).fit(X, Y)
scad_nuc_norm_tuned_with_cv = GlmCV(penalty=NuclearNorm(flavor=NonConvex())).fit(X, Y)