Configure seeds on source level, not on project level

[berekuk]

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"
• imported values can have unexpected correlations; right now each source in project is evaluated with the same seed; demo: https://squigglehub.org/models/quri-squiggle-experiments/diamond-imports-and-seeds-d