[Bugfix][KV Offload] Fix shared-region ownership and startup cleanup

tests/v1/kv_offload/cpu/test_canonical_layout.py

@@ -236,6 +236,7 @@ def canonical_view(rank: int, world_size: int) -> torch.Tensor:
             rank=rank,
             kv_bytes_per_chunk=row_stride,
             cpu_page_size=row_stride // world_size,
+            unlink_owner=len(regions) == writer_tp + reader_tp - 1,
         )
         regions.append(region)
         # The Triton load path dereferences CPU pointers on the GPU, which is

tests/v1/kv_offload/cpu/test_gpu_worker.py

@@ -297,6 +297,7 @@ def test_transfer(
             rank=0,
             kv_bytes_per_chunk=kv_bytes_per_chunk,
             cpu_page_size=cpu_page_size,
+            unlink_owner=True,
         )
 
     worker = CPUOffloadingWorker(