Require model.get_inference_name() and fail loudly on model.name

[corbt]

Summary

• migrate rollout/inference call sites from model.name to model.get_inference_name() across examples/dev scripts/docs
• remove unsuffixed inference alias behavior so trainable model names must include @step
• make TinkerNative return clear 400 errors for missing/invalid @step suffixes
• align LocalBackend server health checks with step-qualified inference names
• update related multi-checkpoint unit coverage

Why

Recent behavior allowed unsuffixed model.name to silently route to stale step-0 adapters in some backends. This change makes incorrect usage fail loudly and standardizes explicit model@step inference routing.

Validation

• uv run --extra backend pytest tests/integration/test_multi_checkpoint_training.py::test_local_backend -q
• uv run --extra tinker pytest tests/integration/test_tinker_native_backend.py -q
• uv run pytest tests/unit/test_multi_checkpoint_inference.py -q -k "served_model_name_uses_base_model_when_lora_enabled or TestStepParsing"
• yes-no-maybe smoke runs for both LocalBackend and TinkerNativeBackend showed eval improvement over run

[arcticfly]

Thanks. Do you want me to update the notebooks as well, or are you already working on that?

[corbt]

@arcticfly if you can that would be great.