Noise Model: assumption without axes will mostly fail

[jdeschamps]

Problem

The assumption that if axes are not given in NoiseModelTrainer.train_from_pairs then axes are either "SC(Z)YX" or "S(Z)YX" has issues:

• will hardly be valid for user's data
• does not allow to unambiguously identify axes, e.g. len("SCYX")==len("SZYX")
• can lead to iteration over the wrong number of channels (e.g. Y)

Axes should probably be specified.