feat(cambricon): enable MLU590 (device map + graph class)

Add the two touchpoints needed to bring up vllm-plugin-FL on Cambricon
MLU (torch_mlu), verified E2E with Qwen3-8B on MLU590:

- VENDOR_DEVICE_MAP: register cambricon -> mlu, so the vendor-device
  lookup no longer raises ValueError for cambricon.
- Graph: select torch.mlu.MLUGraph for device_type == "mlu", so graph
  capture no longer raises NotImplementedError("not support graph").

No dedicated cambricon dispatch backend is required; ops route through
flag_gems (Triton) generically. This is minimal enablement only.

This PR was written in part with the assistance of generative AI.

Author: tengqm

vllm_fl/compilation/graph.py

@@ -51,6 +51,8 @@ class Graph:
         graph = torch.ptpu.PTPUGraph
     elif current_platform.device_type == "gcu":
         graph = torch.gcu.GCUGraph
+    elif current_platform.device_type == "mlu":
+        graph = torch.mlu.MLUGraph
     elif current_platform.device_type == "txda":
         graph = None
     else:

vllm_fl/utils.py

@@ -51,6 +51,8 @@
     "enflame": {"device_type": "gcu", "device_name": "gcu"},
     # Registered backend: vendor/txda
     "tsingmicro": {"device_type": "txda", "device_name": "txda"},
+    # Cambricon MLU (torch_mlu); ops dispatched via flag_gems
+    "cambricon": {"device_type": "mlu", "device_name": "mlu"},
 }