Add option to use same trvl seed for all replicas

[scarlehoff]

Added an option to the runcard (same_trvl_per_replica, I was not sure about the name) so that the training validation split is the same for all replicas.

Closes #1244

[Zaharid]

The name looks clear enough to me.

As an aside,

• I think we should use independent RandomStates instead of mutating the global one. See e.g. https://numpy.org/doc/stable/reference/random/generated/numpy.random.seed.html
• Could we use something like RandomState(trvalseed + replica).randint(0, 2**32) instead of the for loop?