fix(reproducibility): add opt-in strict determinism across trainers

[SoheylM]

Description

Adds an opt-in strict reproducibility path for all training entrypoints while preserving current default behavior.

• Introduces engiopt/reproducibility.py with shared helpers:
  • seed_training(seed)
  • enable_strict_determinism(warn_only=True)
  • make_dataloader_generator(seed)
• Adds strict_determinism: bool = False to all targeted training Args dataclasses.
• Replaces ad-hoc seeding with seed_training(args.seed) in each training script.
• Enables strict deterministic controls only when --strict-determinism is passed.
• Hardens DataLoader reproducibility by supplying seeded generators for shuffle=True loaders.
• Updates README with optional reproducibility usage (--strict-determinism).

Fixes SOH-14 (Linear)

Type of change

• Documentation only change (no code changed)
• Bug fix (non-breaking change which fixes an issue)
• New algorithm (non-breaking change which adds a new model/algorithm)
• Improvement to existing algorithm (non-breaking change which improves functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

Screenshots

N/A

Checklist:

Code Quality

• I have run the pre-commit checks with pre-commit run --all-files
• I have run ruff check . and ruff format
• I have run mypy .
• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings

CleanRL Philosophy (for new/modified algorithms)

• The implementation follows the CleanRL single-file philosophy: all training logic is contained in one file
• The code is reproducible: random seeds are set, PyTorch determinism is enabled
• Hyperparameters are configurable via command-line arguments using tyro
• WandB logging is integrated with --track flag support
• The model can be saved and restored via WandB artifacts (--save-model flag)

Algorithm Completeness (for new algorithms)

• Both training script (algorithm.py) and evaluation script (evaluate_algorithm.py) are provided
• The algorithm works with EngiBench's Problem interface
• The algorithm is added to the README table with correct metadata

Documentation

• I have made corresponding changes to the documentation
• New algorithms include docstrings explaining the approach and any paper references

Validation

• Determinism smoke check (cgan_cnn_2d) run twice on same machine with strict mode and fixed seed; resulting checkpoint tensor hashes matched for both generator and discriminator.
• Non-strict short run completed successfully (default behavior path unchanged).
• Strict mode configured with warn_only=True for nondeterministic ops to warn and continue.

[linear]

SOH-14 Non reproducibility in training algorithms

[g-braeunlich]

Just 2 small suggestions

[SoheylM]

@g-braeunlich quick ping when you have a moment: could you confirm yes/no on the latest TF32 guard update in the open thread? If yes, we’ll resolve and merge. Thanks!