This includes various improvements

• Improved normalization scheme: masking, bfactor sharpening, and downsampling
• PyDantic used for the inputs, so now some parameters are optional (see the example config file)
• Removed the power spectrum normalization
• The SVD is now done in two steps, first each submission gets its own svd, then those eigenvetors are used to compute a common embedding (see paper)
• Plots are now generated automatically, and there is an option to save them.