Improve training/evaluation pipeline for SubRep neural skill generator

[miruts-code]

• Added a real train/val/test split (previously training and evaluation used the same data with no held-out set); split is persisted to a manifest so evaluation always reuses the exact split from training instead of recomputing it.
• Training now runs only on the train split; validation loss is tracked each epoch and used for model selection (best-validation checkpoint, not just the final epoch) and early stopping.
• Replaced hardcoded hyperparameters (batch size, epochs, learning rate, hidden dim, seed, split ratios) with CLI arguments.
• Added mid-training checkpointing, so a crash mid-run doesn't lose all progress.
• Added explicit device selection (GPU if available, else CPU).
• Extended evaluation beyond raw loss: added a certification-focused report covering candidate success/admission rate, rejection reasons, payoff/motive improvement over baseline, comparison against the non-neural candidate policies, and generalization to seeds never used in training -- reusing the existing certification and baseline logic unmodified.
• Updated README.md, generator/README.md, and data/README.md to reflect the above

Description

What does this PR do? (e.g., "Implements 2-head MLP forward pass")

Checklist

• Code follows project style guidelines
• Tests added/updated (tests/test_*.py)
• README updated (if interface changed)
• Verified locally (all tests pass)

Testing Notes

How should the reviewer test this?