Allow for three models using a constructor argument: method_type, which can be "ard", "bnn" or "ridge". By default: "ard".
ARD prunes the weights, making this more resilient to noise.
ARD does not accept weighting, but after using the XGM intensity in the model, the weighting is not very helpful anyway, so removed it by default (although still possible here).
Removed multi output optimization for ARD, since it does not work in this case. Use generic parallelization for ARD and keeping the vectorization of the multi-output meta-estimator only for Ridge.
Adapted plotting scripts for testing and outputting CSV for plots too, in case we want to do plots offline.
Check number of peaks in SPEC spectrum. If it is less than 3, exclude it from training. This pre-selects SPEC data to avoid using low-quality data.