Right now we use the current model's seed for evaluating all imports.
This has a few bad consequences:
imported dists and other random values won't look the same as they looked in the original model
we can't cache outputs and reuse them for imports — "output of this import with its seed" is not the same as "output of this import with the importer's seed"
Right now we use the current model's seed for evaluating all imports.
This has a few bad consequences: