test(cp_shared_kv): load real sgl_kernel first to stop CPU-CI stub leaking

test_cp_shared_kv_runtime installs CPU-CI sgl_kernel stubs at import time via
sys.modules.setdefault + torch.library FRAGMENT defs. On a GPU box, when this
module is collected before the real sgl_kernel loads, the empty stub permanently
shadows it process-wide, so a later real-kernel test (e.g.
fast_topk_transform_ragged_fused in test_nsa_topk_transform) calls a None-returning
lambda and crashes. Importing the real sgl_kernel first makes setdefault keep the
real module (setattr only fills missing names) and the FRAGMENT defs hit the
already-registered path; on CPU-CI the import fails and stubbing proceeds as before.

Verified on GPU: test_cp_shared_kv_runtime alone 120 passed; combined with
test_nsa_topk_transform 125 passed (previously 1 failed / segfault on cleanup).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

test/registered/unit/mem_cache/test_cp_shared_kv_runtime.py

@@ -6,6 +6,20 @@ from unittest.mock import patch
 
 import torch
 
+# Prefer the real compiled sgl_kernel when it is available (e.g. on a GPU box).
+# The CPU-CI stubs below register torch.library ops and a sys.modules module stub
+# at import time. If this module is imported before the real sgl_kernel loads, the
+# empty stub permanently shadows it process-wide, so a later test that needs the
+# genuine kernel (e.g. fast_topk_transform_ragged_fused) calls a None-returning
+# lambda. Loading the real module first makes the FRAGMENT defines hit the
+# already-registered path and keeps sys.modules["sgl_kernel"] pointing at the real
+# module (setattr only fills genuinely-missing names). On CPU-CI the import fails
+# and the stubs are built exactly as before.
+try:
+    import sgl_kernel  # noqa: F401
+except Exception:
+    pass
+
 _sgl_kernel_lib = torch.library.Library("sgl_kernel", "FRAGMENT")