Implement riemannian geometric median

[lkorczowski]

In the rASR matlab implementation, the geometric median is euclidean. For now, python RASR uses also the euclidean geometric median (with the different optimization method though).

We might want to implement a riemannian Geometric Median.

Advantage:

• Fisher metric instead of euclidean metric is better suited for covariance matrices
• therefore we might expect better overall performance of the rASR

Disadvantage:

• doesn't handle non-positive definite matrices (but we could compensate by using robust cov estimator or regularization)
• slower (but it should be a problem as the median is only required during calibration)
• we need to find a way to optimize the cost function (e.g. we can compute analytically the gradien, use pymanopt or with a handmade gradien descent)

Right-now, the advantage is not clear so it is not a priority.

[lkorczowski]

Florent Bouchard proposed us to give the gradien and hessian so we can find the riemannian geometric median using pymanopt TrustRegions() and/or we can simply ask autograd to find them (the cost function is very simple).

[qbarthelemy]

The Riemannian geometric median is available in pyRiemann since v0.4 https://pyriemann.readthedocs.io/en/latest/generated/pyriemann.utils.median_riemann.html