fix: 修复推理引擎 batch decode 中多项正确性与并发问题

- scheduler: decode 分组由幂次分桶改为精确 next_pos，消除 KV cache 位置错乱
- task: activate() 加锁操作 active_tasks，消除数据竞争
- engine: wait_completion 加超时，防止分配失败时永久死锁
- sample: TopKStrategy 向量化为 per-sample threshold，尊重各 task 的 top_k
- cache: Storage.write/gather 中 -1 页改用 mask 处理，防数据污染
- executor: prefill 逐 task 循环改为单次 tensor 调用

tests/inference/test_sample.py

@@ -48,12 +48,12 @@ def test_top_k_skip_when_zero():
 
 
 def test_top_k_batch_tensor():
-    """When top_k is a batch tensor, max element governs k for all rows."""
+    """Each row respects its own top_k."""
     logits = torch.tensor([[0.1, 0.5, 0.3], [0.9, 0.2, 0.1]])
     s = TopKStrategy(top_k=torch.tensor([2, 1]))
     result = s.apply(logits.clone(), filter_value=-1e9)
     assert (result[0] > -1e9).sum() == 2
-    assert (result[1] > -1e9).sum() == 2
+    assert (result[1] > -1e9).sum() == 1
 
 
 def test_top_p_nucleus_filtering():