run_eval: forward --device to winml build subprocess

[vortex-captain]

Summary

winml config, winml perf, and winml eval already receive --device from run_eval; winml build was the only step missing it. With this change, the build subprocess respects the caller's device choice instead of relying on its own auto-detection.

Companion to #667 (which forwarded --ep and added device-aware EP auto-select in the build CLI). With both PRs in, run_eval consistently propagates --device/--ep to every winml subprocess it spawns.

Diff

        build_args = [
            *WINML_CLI,
            "build",
            "-c",
            str(sub_cfg),
            "-m",
            entry.hf_id,
            "--use-cache",
+            \"--device\",
+            device,
        ]
        if ep:
            build_args += [\"--ep\", ep]

Verification

E2E on a cold cache:

```
python scripts/e2e_eval/run_eval.py --hf-model BAAI/bge-base-en-v1.5 --ep cpu --device cpu --eval-type perf
```

→ exit 0, `[PASS] 2.4s`. Generated build config has `"compile": null` (correct for cpu — no EP-specific compile needed).