[FEP-0062] Cambricon MLU GRPO E2E acceptance tracking

## Summary

Tracking issue for the **Cambricon MLU** GRPO E2E acceptance run of [FEP-0062](https://github.com/flagos-ai/community/pull/62) (verl-hardware-plugin v0.1.0).

Current v0.1.0 acceptance status (per FEP-0062 Test Plan — Qwen3-0.6B / GSM8K / FSDP / vLLM rollout):

| Platform | E2E GRPO run |
|----------|--------------|
| MetaX | ✅ passed |
| Iluvatar | ✅ passed |
| Cambricon MLU | ❌ failed — this issue |

## Scope

- Root-cause the Cambricon MLU failure (plugin side: CNCL init / Ray `MLU` resource / `torch_mlu` adaptation — vs. vendor stack: driver / firmware / vendor torch).
- Land the fix in [verl-hardware-plugin](https://github.com/verl-project/verl-hardware-plugin) and re-run the E2E acceptance per the FEP-0062 Test Plan.
- If the re-run passes within the 2.2 rc window (bugfixes accepted until **2026-09-24**, see [#47](https://github.com/flagos-ai/community/issues/47)), FEP-0062 flips to `Implemented` with all three chips at GA. Otherwise this item retargets the next release.

## Test results / 测试结果

@Al2O3Al2O3 : please attach the failure details here — environment (driver / vendor torch / CNCL versions), reproduction command, logs, and subsequent re-run results._

## References

- FEP document: [`fep/sig-framework/0062-verl-hardware-plugin-v0.1.0-multi-chip-grpo-e2e.md`](https://github.com/flagos-ai/community/blob/main/fep/sig-framework/0062-verl-hardware-plugin-v0.1.0-multi-chip-grpo-e2e.md)
- Plugin repository: https://github.com/verl-project/verl-hardware-plugin
- Release tracking: #47