Optimize torchair kv_consumer padding logic (#3526)

### What this PR does / why we need it?
Optimize torchair kv_consumer padding logic. Only pad when it is spec
decoding

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

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

Signed-off-by: xuyexiong <xuyexiong@huawei.com>

vllm_ascend/torchair/torchair_model_runner.py

@@ -85,7 +85,8 @@ class NPUTorchairModelRunner(NPUModelRunner):
     def _may_pad_kv_consumer_num_seq(self):
         # pd disaggregation scenario need redundant_batch_sizes to avoid each batch's seq_len exceed 16 tokens
         # self.max_num_reqs here is greater than the actual maximum request number
-        if self.is_kv_consumer:
+        if self.decode_token_per_req > 1 and self.is_kv_consumer:
+            # applied only when speculative decoding is active
             FIA_SEQ_LEN_LIMIT = 16
             new_max_num_reqs = self.max_num_reqs + math.ceil(
                 self.max_num_reqs / FIA_SEQ_LEN_LIMIT) + math.ceil(