[Perf][PrefixCache] Avoid per-step blocking write in OmniTensorPrefix… (#4106)

Signed-off-by: LHXuuu <xulianhao.xlh@antgroup.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

examples/online_serving/qwen3_omni/qwen3_omni_moe_thinking.yaml

@@ -20,7 +20,7 @@ stage_args:
       trust_remote_code: true
       engine_output_type: text
       distributed_executor_backend: "mp"
-      enable_prefix_caching: false
+      enable_prefix_caching: true
       hf_config_name: thinker_config
       tensor_parallel_size: 2
     final_output: true