Add random seed support for reproducible workflow runs

[dtch1997]

Problem

Currently, when running replications in a parameter sweep (e.g., with suffixes rep-1, rep-2, rep-3), there's no way to control the random seed to ensure reproducibility or to intentionally vary random initialization across replications.

Proposed Solution

Add support for specifying random seeds in workflow configurations, likely in the submit_training config:

submit_training=SubmitTrainingConfig(
    model="meta-llama/Llama-3.2-1B",
    hyperparameters={
        "seed": 42,  # or random_seed
        ...
    },
    ...
)

This would allow:

1. Exact reproducibility when using the same seed
2. Controlled variation across replications by sweeping over different seeds
3. Better experimental control

Example Use Case

param_grid = {
    "submit_training.hyperparameters.learning_rate": [1e-4, 5e-5, 1e-5],
    "submit_training.hyperparameters.seed": [42, 43, 44],  # 3 seeds for replications
}