[Kernel] Rewrite JIT custom all-reduce (v2) with a decoupled kernel/storage design (#31049)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: root <root@GPUC5A6.maas>

python/sglang/jit_kernel/tests/utils.py

@@ -14,7 +14,7 @@ def multigpu_pytest_main(
     pre_launch_fn: Optional[Callable[[List[int]], None]] = None,
     timeout: Optional[int] = 600,
 ) -> None:
-    """cudalib-style multi-GPU pytest entry point.
+    """Torchrun-based multi-GPU pytest entry point.
 
     Drop this at the bottom of a test file::