[BugFix]Fix FIA input err in DSv3.1 (#5059)

### What this PR does / why we need it?
When use mtp, full decdoe only and async_scheduling together, finding a
input err for FIA ops due to the non-increasing input
of the 'actual_seq_lengths'. This bug is caused by the filling the
variable ‘query_start_loc’. We need to fill the query_start_loc' s end
by the 'cu_num_tokens' instead of '-1'

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

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: hust17yixuan <303660421@qq.com>

vllm_ascend/worker/model_runner_v1.py

@@ -769,6 +769,7 @@ class NPUModelRunner(GPUModelRunner):
 
         self.query_start_loc.np[0] = 0
         self.query_start_loc.np[1:num_reqs + 1] = cu_num_tokens
+        self.query_start_loc.np[num_reqs + 1:].fill(cu_num_tokens[-1])
         self.query_start_loc.copy_to_gpu()
 
         self.seq_lens.np[:num_reqs] = (
@@ -776,8 +777,6 @@ class NPUModelRunner(GPUModelRunner):
             num_scheduled_tokens)
         self.seq_lens.copy_to_gpu()
 
-        # Fill unused with -1. Needed for reshape_and_cache
-        self.query_start_loc.gpu[num_reqs + 1:].fill_(-1)
         self.seq_lens.gpu[num_reqs:].fill_(0)
 
         self.query_lens = torch.from_numpy(num_scheduled_tokens)