Implement classification with NCH (Nearest Convex Hull)

[toncho11]

Description:
In NCH, the minimal distance between a tested matrix and the convex hull of a set of matrices is estimated. This hull is the actual convex hull of the set of matrices (corresponding to each class). There is no training. Calculating the distance to the hull is an optimization problem and it is calculated for each testing sample (testing SPD matrix).

Implementation:
Add a new class "QuantumNCH" in classyfication.py which will use distance_logeuclid_cpm. distance_logeuclid_cpm is the new distance to hull that we need for this classifier and it is implemented (prototype) in #244. We will probably first start with non-quantum optimizer. The fit() method will simply store the data, so that it can be used in the predict() with distance_logeuclid_cpm.

Prediction
A distance is calculated to each hull (one hull per class). The lower distance selects the predicted class.

Notes:
The fact that we perform an optimization on each test sample and for each candidate class means that the classification (prediction) task will be slow in general. Also the NCH algorithm might be more susceptible to outliers in the data.