Add non-linear 2-state decay classifier

[MarcSerraPeralta]

This is an improvement over the linear 2-state decay classifier:

1D decay classifier

2D decay classifier

[MarcSerraPeralta]

The general covariance matrix can be compted from $G = R(\theta)diag(\sigma_0^2, \sigma_1^2) R(-\theta)$ where $R(\theta)$ is the rotation matrix for 2D for an angle $\theta$.

Then, the expression for $G$ is:

G = [[sigma0**2 * np.cos(theta)**2 + sigma1**2 * np.sin(theta)**2, (sigma0**2 - sigma1**2) * np.sin(theta) * np.cos(theta)  ],
     [(sigma0**2 - sigma1**2) * np.sin(theta) * np.cos(theta)    , sigma0**2 * np.sin(theta)**2 + sigma1**2 * np.cos(theta)**2]]

[MarcSerraPeralta]

The number of free parameters in this classifier can be very large, in the worse case: 2 * 8 = 16.
For each of the two states, we have 8 parameters: mu_0_x, mu_0_y, mu_1_x, mu_1_y, sigma0, sigma1, theta, amplitude (where amplitude corresponds to the amount of mixture of the two gaussians, and theta corresponds to the general covariance matrix). I have chosen different means for the two states because in the pictures from the experimental data (attached in the messages of this issue), they do not seem to match.

Note that currently, the maximum number of free parameters in a classifier is 9, so this would be almost doubling them...