Fix (sys.meta_path is None, Python is likely shutting down) (#116)

* Fix (sys.meta_path is None, Python is likely shutting down).

* Add warning.

streaming/base/dataset.py

@@ -162,8 +162,14 @@ def __init__(self,
         # Partition state.
         self._partition_state = None
 
+        # Initialize the World context.
+        #
+        # Beware: This information is for the per-rank process. DataLoader worker processes may see
+        # different values for these fields. We are saving the rank World here because we cannot
+        # instantiate a World inside the StreamingDataset destructor.
+        self._rank_world = world = World()
+
         # Seed is set below.
-        world = World()
         self.num_canonical_nodes = num_canonical_nodes
         self.batch_size = batch_size
         self.shuffle_seed = shuffle_seed
@@ -779,7 +785,7 @@ def _cleanup_shared_memory(self, shm: Any, world: World) -> None:
 
     def __del__(self):
         # Wait for the local rank 0 process
-        world = World()
+        world = self._rank_world
         wait_for_local_leader(world)
 
         # Clean up shared memory resources