Fix Jev `test_deciders` OOM and `test_jev_9b` expected values
- `test_deciders`: the deciders are 27B checkpoints (~54 GB in bfloat16), loaded
with `device_map="auto"` so they offload to the CPU instead of running out of
accelerator memory. `check_decision` passes `pipeline_class=DecisionPipeline`
so `vllm-sr/d3`, whose config declares its own remote `decision` pipeline,
doesn't require `trust_remote_code=True`.
- `test_jev_9b`: update the expected values; bfloat16 and float16 on GPU and
float32 on CPU all agree with the new ones.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>