This allows us to use a much smaller set of markers (and thus save a little bit of time for the recomputation) instead of always taking the union of all markers from train_single().
Users get more control over how they define the "universal" markers for each label within each reference, which could be different from the pairwise markers.
Also allows us to remove a dependency on train_single.hpp in the integrated training/classification code.
This allows us to use a much smaller set of markers (and thus save a little bit of time for the recomputation) instead of always taking the union of all markers from
train_single().Users get more control over how they define the "universal" markers for each label within each reference, which could be different from the pairwise markers.
Also allows us to remove a dependency on
train_single.hppin the integrated training/classification code.