feat(cambricon): blacklist flag_gems empty via built-in config

Add vllm_fl/dispatch/config/cambricon.yaml so the `empty` operator
blacklist is applied automatically on Cambricon MLU, instead of
requiring the VLLM_FL_FLAGOS_BLACKLIST env var at launch.

flag_gems `empty` derives its Triton launch grid from the element
count (e.g. 607744 for the Qwen3 vocab embedding), which exceeds the
MLU Triton grid limit of 65535 and raises OutOfResources. The config
falls the op back to torch_mlu.

The file is named `cambricon.yaml` to match current_platform.vendor_name
("cambricon"), which is how get_config_path() resolves the platform
config (Priority 3). Verified E2E on MLU590: serve reaches startup
complete and inference returns coherent output with no env var set.

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

Author: tengqm

vllm_fl/dispatch/config/cambricon.yaml

@@ -0,0 +1,20 @@
+# vLLM-FL Dispatch Configuration for Cambricon MLU (torch_mlu)
+# Auto-loaded when running on Cambricon MLU hardware (vendor_name == "cambricon").
+#
+# There is no dedicated vendor/cambricon dispatch backend; ops route through
+# FlagOS (flag_gems, Triton). Only per-op exceptions are listed below.
+
+# Preferred default backend type: flagos, vendor, reference
+prefer: flagos
+
+# Strict Mode:
+#   true  = Raise an error immediately on failure; do not attempt other backends.
+#   false = Attempt the next available backend in sequence upon failure (Default).
+strict: false
+
+# FlagOS operator blacklist
+flagos_blacklist:
+  # flag_gems `empty` computes a Triton launch grid from the element count
+  # (e.g. 607744 for the Qwen3 vocab embedding), which exceeds the MLU Triton
+  # grid limit of 65535 and raises OutOfResources. Fall back to torch_mlu.
+  - empty