Improve BERT-like models performance with better self attention (#9124)

* Improve BERT-like models attention layers

* Apply style

* Put back error raising instead of assert

* Update template

* Fix copies

* Apply raising valueerror in MPNet

* Restore the copy check for the Intermediate layer in Longformer

* Update longformer

src/transformers/dependency_versions_table.py

@@ -40,8 +40,8 @@ deps = {
     "sphinx-rtd-theme": "sphinx-rtd-theme==0.4.3",
     "sphinx": "sphinx==3.2.1",
     "starlette": "starlette",
-    "tensorflow-cpu": "tensorflow-cpu>=2.0",
-    "tensorflow": "tensorflow>=2.0",
+    "tensorflow-cpu": "tensorflow-cpu>=2.3",
+    "tensorflow": "tensorflow>=2.3",
     "timeout-decorator": "timeout-decorator",
     "tokenizers": "tokenizers==0.9.4",
     "torch": "torch>=1.0",