[P/D] force with_prefill true after allreduce in kv producer (#3768)

### What this PR does / why we need it?
force with_prefill true after allreduce in kv producer

- vLLM version: v0.11.0rc3
- vLLM main:
c9461e05a4

---------

Signed-off-by: liziyu <liziyu16@huawei.com>

vllm_ascend/distributed/mooncake_layerwise_connector.py

@@ -1151,7 +1151,7 @@ class MooncakeLayerwiseConnectorWorker:
                       connector_metadata: MooncakeLayerwiseConnectorMetadata,
                       **kwargs) -> None:
         """MooncakeLayerwiseConnector does not save explicitly."""
-        if self.kv_role == 'kv_producer':
+        if self.kv_role == 'kv_producer' and connector_metadata.request.keys():
             if self.pd_head_ratio != 1:
                 if self.current_layer != 0:
                     self.completion_event.wait()