Random number generator seeding

While MaggotUBA-based taggers automatically seed random number generators on the Python side, for bit-scale reproducibility, the sampling step of the training dataset does not offer a similar control on Julia-side random number generators.

Until recently, the sampling step was performed independently. As it is now an integral part of the standard workflow, and functional tests are getting automated, we need a --seed argument for the train command so that the entire training procedure can generate equal files ("equal" in the meaning of e.g. the cmp command).