feat(training): port lerobot_act_train_end_to_end AzureML pipeline

[junkataoka]

Problem

The toolchain trains LeRobot policies via a single AzureML CommandJob (training/il/scripts/submit-azureml-lerobot-training.sh) and stops there. There is no AzureML Pipeline that bundles preprocess → train → evaluate as a DAG, and no opt-in register step. Users have to manually chain three submits and check artifacts between each.

Proposed solution

Port lerobot_act_train_end_to_end as an AzureML Pipeline

• 3-step DAG (preprocess → train → evaluate) as the default pipeline
• 4-step DAG (+ register) as an opt-in second pipeline file
• No eval-gates-register enforcement
• No external lineage hooks / no eval-owned candidate-tag patching
• Compute names parameterized as pipeline inputs
• Hydra removed from train / evaluate components in favor of env-var style consistent with the rest of the toolchain

Acceptance criteria

• az ml job create --file training/il/workflows/azureml/lerobot-pipeline.yaml runs the 3-step DAG end-to-end
• Opt-in 4-step pipeline registers the trained model on success
• No engagement-specific args remain in component YAMLs
• README documents both pipelines and submit script

Related

• PR: feat(training): port lerobot_act_train_end_to_end as azureml pipeline #927 (draft)