[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>

vllm_ascend/utils.py

@@ -51,6 +51,7 @@ ACL_FORMAT_FRACTAL_NZ = 29
 _CUSTOM_OP_ENABLED = None
 _CURRENT_STREAM = None
 _PREFETCH_STREAM = None
+_GLOBAL_STREAM = None
 _SHARED_EXPERTS_CALCULATION_STREAM = None
 _ASCEND_CUSTOMOP_IS_REIGISTERED = False
 _DEFAULT_BUFFER_SIZE = 200
@@ -292,6 +293,15 @@ def prefetch_stream() -> torch.npu.Stream:
     return _PREFETCH_STREAM
 
 
+def global_stream() -> torch.npu.Stream:
+    global _GLOBAL_STREAM
+    if _GLOBAL_STREAM is None:
+        # when this function is called before any stream is set,
+        # we return the default stream.
+        _GLOBAL_STREAM = torch_npu.npu.Stream()
+    return _GLOBAL_STREAM
+
+
 def shared_experts_calculation_stream() -> torch.npu.Stream:
     global _SHARED_EXPERTS_CALCULATION_STREAM
     if _SHARED_EXPERTS_CALCULATION_STREAM is None: