Selecting the type of embedding for tractogram

[Paolopost]

Currently the choice of embedding is hardcoded in the loading procedure. By defaults is set to "dismatrix", e.g. dissimilarity representation. We should support the choice of different embeddings. Some trials are carried out with embedding obtained from convolutional autoencoders, such as FINTA and ECLIPSE. We should consider whether to adopt a default embedding based on resampling to a fixed number of points (plus flattening of coordinates).