feat(evaluation): add pi0 dispatch and VLA schema v1 reconciliation to policy_runner

[junkataoka]

Problem

evaluation/sil/policy_runner.py only dispatches ACT policies. With pi0 training landing (#916), there is no SIL eval path for pi0 checkpoints. Additionally, P1 (#927) introduces VLA evaluation schema v1 emission in run_evaluation.py, which needs to be reconciled with the existing toolchain metrics schema for consistency across policies.

Proposed solution

• Add pi0 / pi0_fast dispatch branch to policy_runner.py
• Add a pi0 eval submit script + workflow YAML
• Reconcile VLA schema v1 fields with the toolchain's existing eval output so both ACT and pi0 evals emit the same shape

Acceptance criteria

• policy_runner.py accepts pi0 checkpoints and runs SIL rollouts
• Eval submit script + workflow YAML for pi0
• ACT and pi0 emit identical schema v1 JSON
• Metric schema documented in eval README

Dependencies

• Blocked on: feat(training): port lerobot_act_train_end_to_end as azureml pipeline #927 (schema v1 emitter contract must stabilize in review first)

Estimate

3–4 person-days