WIP: Correlation matrix

[pierrepaleo]

This MR brings the following features:

• Build the correlation matrix ("two-times correlation function"). It can be the full matrix of shape (n_frames, n_frames), or a reduced matrix of shape (n_frames, n_times).
• From the correlation matrix, get the standard deviation ("error bars")

In a future MR, the above should also be done "on-line", i.e update an existing correlation matrix with a new series of frames.

Calculating the standard deviation without building the correlation matrix is out of scope (more complicated!).

The correlation matrix can be built in two ways:

• Directly from the "space-sparse" format, i.e data compacted frame per frame (output of future acquisition system). The class SMatrixEventCorrelator has three methods (build_correlation_matrix_v1, v2, v3) corresponding to three approaches (and OpenCL kernels). v3 is the fastest.
• From the "time-sparse" format, i.e data compacted along the time axis. For "really" sparse data, it is much more efficient. Converting from space-compacted to times-compacted can be done quite quickly.