[0.18.0][BugFix] Update capture sizes after rounding operations (#8380)

### What this PR does / why we need it?
This PR is partially cherry-picked from #8172.

This PR aims to fix mismatched capture sizes after rounding operations
when using sp or speculative. The reason is that original
`self.cudagraph_capture_sizes` is no longer updated and remains as the
initial sizes. Now we use `self.cudagraph_dispatcher.get_capture_descs`
to the get up-to-date sizes.

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
by ci

Signed-off-by: Zetong Li <slippersss@126.com>

vllm_ascend/worker/model_runner_v1.py

@@ -3298,12 +3298,19 @@ class NPUModelRunner(GPUModelRunner):
         with update_pass_config(self):
             super()._check_and_update_cudagraph_mode(attention_backends, kv_cache_groups)
 
+        capture_descs = self.cudagraph_dispatcher.get_capture_descs()
+        capture_sizes = sorted({
+            desc.num_tokens
+            for _, descs in capture_descs
+            for desc in descs
+        })
+
         # NOTE: Since aclgraph_batch_sizes cannot be determined until here,
         # we set the graph params right before initializing the keys.
         if self.use_aclgraph:
-            set_graph_params(self.cudagraph_batch_sizes)
+            set_graph_params(capture_sizes)
             if self.speculative_config:
-                set_draft_graph_params(self.cudagraph_batch_sizes)
+                set_draft_graph_params(capture_sizes)
 
     def capture_model(self) -> None:
         gpu_model_runner_cls = next((cls for cls in self.__class__.__mro__ if cls.__name__ == "GPUModelRunner"), None)