Add shuffle seed to MAGIC

[SrGonao]

Separates initialization and data order from MAGIC training, by adding a new seed to the config.

[norabelrose]

I'm not totally sure this decoupling is needed— did you find that you needed it in your experiments? Also, the doc string for seed in this PR seems misleading. It's only getting used to create the subsets for leave-k-out validation and nothing else. If we're going to have two seeds, both of them should have more descriptive names.

[luciaquirke]

We were investigating the impact of data order on the results. Reviewing this it seems like we could do this without additional seeds by pipelining the magic and validate commands like [pseudocode]:

magic:
  run_path: 0
  seed: 0 
  skip_validate: True
magic:
  run_path: 1
  seed: 1
  skip_validate: True
validate:
  run_path: 0
  seed: 0
validate:
  run_path: 1
  seed: 0

@norabelrose is that what you were thinking of?