Update ML Decoder

[fffffgggg54]

Update ML Decoder's TransformerDecoderLayerOptimal module to comply with what nn.TransformerDecoder expects. Current changes work with resnet50.

add_ml_decoder_head needs to be updated for other models. In my limited testing, the following case works with RegNet:

elif hasattr(model, 'head'):    # ClassifierHead and ConvNext
    if hasattr(model.head, 'flatten'):  # ConvNext case
        model.head.flatten = nn.Identity()
    model.head.global_pool = nn.Identity()
    del model.head.fc
    num_classes = model.num_classes
    num_features = model.num_features
    model.head.fc = MLDecoder(num_classes=num_classes, initial_num_features=num_features)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.