[main][bugfix] bugfix for minicpm models (#3527)

### What this PR does / why we need it?
bugfix for minicpm-2b and minicpm3-4b

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: Wang Kunpeng <1289706727@qq.com>

vllm_ascend/worker/model_runner_v1.py

@@ -1346,6 +1346,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
         positions_cpu = self.positions_cpu[:num_input_tokens]
         positions = self.positions[:num_input_tokens]
         seq_lens_cpu = self.seq_lens_cpu[:num_reqs]
+        attn_state = self._build_attn_state(num_reqs, num_scheduled_tokens,
+                                            num_valid_tokens)
         self.attn_mask = self._make_attention_mask(seq_lens=seq_lens_cpu,
                                                    position=positions_cpu,
                                                    attn_state=attn_state)