Explore problems where our algorithm can outperform the rest

[theogf]

• High-dimensional Gaussian like targets (see #3)
• Problems where inverting the covariance matrix is the bottleneck

[theogf]

Just an idea : Should we actually consider an experiment for Gaussian Processes with a very high number of points, where classical inference is not possible, i.e. > 1e3 samples. And see how our algorithm performs, and compare it with other scaling approach like inducing points and co

[theogf]

Tried this in #5 and it does not work so well, one still need to differentiate through all the samples and it scales quadratically with the number of points