Add a configuration option to retain all intermediate output files from predictors. Currently, the pipeline only uses a subset of results (e.g., affinity from netMHCpan), but saving everything (such as rank, is binder...) would make it much easier to reuse these results later without rerunning these predictors.