[v0.18.0][Bugfix][EAGLE] Fix FIA pad bug under max concurrency (#7754)

cherry picked from https://github.com/vllm-project/vllm-ascend/pull/7740
Fixes padding problems of FIA op under max concurrency.

- vLLM version: v0.18.0
- vLLM main:
35141a7eed

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

vllm_ascend/spec_decode/eagle_proposer.py

@@ -168,6 +168,8 @@ class SpecDecodeBaseProposer(EagleProposer):
             # RoPE need (max_num_tokens,)
             self.positions = torch.zeros(self.max_num_tokens, dtype=torch.int32, device=device)
 
+        self.token_arange_np = np.arange(self.max_num_tokens + 1)
+
     def _get_model(self) -> nn.Module:
         """
         Default method to call get_model(). Can be overridden by subclasses which