[BUG] Clear _job_queue in parallel_imap for tests (#204)

If compilation with `nvcc` failed `_job_queue` might be not empty caused followed tests will fail with "Cannot call parallel_map recursively."

Clear `_job_queue` in `parallel_imap` before every test. 


Tested manually. 
Before: subsequent tests failed with "Cannot call parallel_map recursively."
After: only one test failed.

tests/conftest.py

@@ -67,14 +67,17 @@ def pytest_collection_modifyitems(config, items):
 
 
 @pytest.fixture(autouse=True)
-def clear_memory_cache():
+def clear_before_test():
     """
     Clear the memory cache before each test.
     """
     # run before each test
     import gc
     import torch
 
+    # If previous test failed job queue for `paralles_imap` might be not free
+    hidet.utils.multiprocess._job_queue = None
+
     torch.cuda.empty_cache()
     hidet.runtime.storage.current_memory_pool('cuda').clear()
     gc.collect()  # release resources with circular references but are unreachable