fix: close envs between tasks to prevent worker process accumulation

eval_policy_all never closed environments after each task completed,
causing AsyncVectorEnv worker processes to accumulate (N_tasks × n_envs).
This led to OOM, BrokenPipeError and EOFError on multi-task benchmarks.

Also fixes:
- AsyncVectorEnv compat in envs/utils.py (use get_attr/call instead of .envs)
- Tuple task handling in tokenizer_processor and lerobot_eval
- _LazyAsyncVectorEnv for deferred worker spawning in LIBERO

Made-with: Cursor

tests/processor/test_tokenizer_processor.py

@@ -189,6 +189,30 @@ def test_list_of_strings_tokenization(mock_auto_tokenizer):
     assert attention_mask.shape == (2, 8)
 
 
+@require_package("transformers")
+@patch("lerobot.processor.tokenizer_processor.AutoTokenizer")
+def test_tuple_of_strings_tokenization(mock_auto_tokenizer):
+    """Test tokenization of a tuple of strings (returned by VectorEnv.call())."""
+    mock_tokenizer = MockTokenizer(vocab_size=100)
+    mock_auto_tokenizer.from_pretrained.return_value = mock_tokenizer
+
+    processor = TokenizerProcessorStep(tokenizer_name="test-tokenizer", max_length=8)
+
+    transition = create_transition(
+        observation={"state": torch.tensor([1.0, 2.0])},
+        action=torch.tensor([0.1, 0.2]),
+        complementary_data={"task": ("pick up cube", "place on table")},
+    )
+
+    result = processor(transition)
+
+    observation = result[TransitionKey.OBSERVATION]
+    tokens = observation[f"{OBS_LANGUAGE}.tokens"]
+    attention_mask = observation[f"{OBS_LANGUAGE}.attention_mask"]
+    assert tokens.shape == (2, 8)
+    assert attention_mask.shape == (2, 8)
+
+
 @require_package("transformers")
 @patch("lerobot.processor.tokenizer_processor.AutoTokenizer")
 def test_custom_keys(mock_auto_tokenizer):