Add support of MViTv2 video variants (#6373)

* Extending to support MViTv2

* Fix docs, mypy and linter

* Refactor the relative positional code.

* Code refactoring.

* Rename vars.

* Update docs.

* Replace assert with exception.

* Updat docs.

* Minor refactoring.

* Remove the square input limitation.

* Moving methods around.

* Modify the shortcut in the attention layer.

* Add ported weights.

* Introduce a `residual_cls` config on the attention layer.

* Make the patch_embed kernel/padding/stride configurable.

* Apply changes from code-review.

* Remove stale todo.

docs/source/models/video_mvit.rst

@@ -12,7 +12,7 @@ The MViT model is based on the
 Model builders
 --------------
 
-The following model builders can be used to instantiate a MViT model, with or
+The following model builders can be used to instantiate a MViT v1 or v2 model, with or
 without pre-trained weights. All the model builders internally rely on the
 ``torchvision.models.video.MViT`` base class. Please refer to the `source
 code
@@ -24,3 +24,4 @@ more details about this class.
     :template: function.rst
 
     mvit_v1_b
+    mvit_v2_s

test/test_models.py

@@ -309,6 +309,9 @@ _model_params = {
     "mvit_v1_b": {
         "input_shape": (1, 3, 16, 224, 224),
     },
+    "mvit_v2_s": {
+        "input_shape": (1, 3, 16, 224, 224),
+    },
 }
 # speeding up slow models:
 slow_models = [