[Performance] Pre-issued exponential distribution operator. (#4908)

Pre-issued exponential distribution operator.

Result:
Single inference saves 200-300 microseconds.
before：

<img width="2257" height="1058" alt="2"
src="https://github.com/user-attachments/assets/c1da19e2-a439-42cb-9d7c-c0218e61fd4c"
/>

After：

<img width="2211" height="342" alt="image"
src="https://github.com/user-attachments/assets/03c84292-c802-4755-949c-4266a9a72fc0"
/>


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

---------

Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>

tests/ut/sample/test_sampler.py

@@ -17,9 +17,12 @@ class TestAscendSampler(TestBase):
 
 class TestAscendTopKTopPSampler(TestBase):
 
+    @mock.patch("vllm_ascend.sample.sampler.random_sample")
     @mock.patch("torch_npu.npu_top_k_top_p")
-    def test_npu_topk_topp_called_when_optimized(self, mock_npu_op):
+    def test_npu_topk_topp_called_when_optimized(self, mock_npu_op,
+                                                 mock_random_sample):
         mock_npu_op.return_value = (torch.randn(1, 3))
+        mock_random_sample.return_value = torch.randn(3)
         sampler = AscendTopKTopPSampler()
 
         logits = torch.tensor([[1.0, 2.0, 3.0]])