[v0.18.0][BugFix] Fix Qwen3.5 MoE FC1 error under high concurrency when dp>1 (#8395)

### What this PR does / why we need it?
GDN Attention uses FIA's query_start_loc (padded), which may cause
conv1d update errors under high concurrency when dp > 1, and this PR is
to make GDN use its own query_start_loc (unpadded).

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
- vLLM version: v0.18.0

Signed-off-by: Wangbingjie <wangbj1207@126.com>

vllm_ascend/worker/model_runner_v1.py

@@ -2201,7 +2201,7 @@ class NPUModelRunner(GPUModelRunner):
             if self._has_gdn:
                 attn_group = self.attn_groups[kv_cache_gid][0]
                 builder = attn_group.get_metadata_builder(0)
-                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
+                if isinstance(builder, GDNAttentionMetadataBuilder):
                     cm.query_start_loc_cpu = self.gdn_query_start_loc.cpu[: num_reqs_padded + 1]
                     cm.query_start_loc = self.gdn_query_start_loc.gpu[: num_reqs_padded + 1]
 
@@ -2399,6 +2399,9 @@ class NPUModelRunner(GPUModelRunner):
             num_reqs_padded = self._pad_query_start_loc_for_fia(
                 num_tokens_padded, num_reqs_padded, num_reqs, cudagraph_runtime_mode, batch_desc.num_reqs
             )
+            if self._has_gdn:
+                self.gdn_query_start_loc.np[1 : num_reqs_padded + 1] = cum_num_tokens
+                self.gdn_query_start_loc.copy_to_gpu()
 
             pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
             attn_metadata, _ = self._build_attention_metadata(