Fixing a bug where long tokenizer files would be incorrectly deserialized (#459)

* Fixing a bug where long tokenizer files would be incorrectly
deserialized

- Add a bunch of tests to check deserialization behaviour
- One tests also confirms current Single deserialization of Sequence.

* Better test locations for Windows + no file dependency in Python binding
Rust side.

* Adressing @n1t0 comments.

bindings/python/tests/utils.py

@@ -75,6 +75,13 @@ def train_files(data_dir):
     }
 
 
+@pytest.fixture(scope="session")
+def albert_base(data_dir):
+    return download(
+        "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v1-tokenizer.json"
+    )
+
+
 def multiprocessing_with_parallelism(tokenizer, enabled: bool):
     """
     This helper can be used to test that disabling parallelism avoids dead locks when the