Consider migrating quickstart examples to OpenWeights backend

[dtch1997]

Not everyone has Tinker API key

Also OpenWeights has a lot more degrees of freedom in terms of what you can do

The downside is that it'll take much longer - maybe closer to ~1h instead of <1 min

Conclusion: Maybe we should just add OpenWeights examples instead of modifying the current Tinker examples

[coderabbitai]

A summary of the changes CodeRabbit can apply:

• Migrate quickstart examples from Tinker to OpenWeights by adding "openweights" support and API-key fallback, updating environment template, docs, README, examples and configs, and adding a migration guide — modify motools/steps/submit_training.py (add backend_name validator), motools/training/backends/openweights.py (OPENWEIGHTS_API_KEY support), .env.template (OPENWEIGHTS_API_KEY), docs/quickstart.md, README.md, docs/demo.md (backend/model/timing updates), docs/MIGRATION_OPENWEIGHTS.md (new), examples/hello_world_minimal.py (backend class, model, hyperparameter names), and examples/configs/train_evaluate_hello_world.yaml (backend_name, model, hyperparameter names).

• Migrate MOTools training from Tinker to OpenWeights by adding OPENWEIGHTS_API_KEY to .env.template, updating README, docs/quickstart.md, docs/demo.md, examples (hello_world_minimal.py, examples/configs/train_evaluate_hello_world.yaml, examples/hello_world.py references), adding a migration guide docs/MIGRATION_OPENWEIGHTS.md, changing backend imports/usages to OpenWeightsTrainingBackend and model/hyperparameter names (unsloth/Llama-3.2-1B, epochs/r), and updating motools/training/backends/openweights.py to read OPENWEIGHTS_API_KEY from the environment.

• ✅ Create PR with these edits
• 📋 Get copyable edits