[Feature] Add MXNet2.0 Compatible Examples

[barry-jin]

Currently, most of the examples in master branch are removed in #20602 because of compatibility issues. These can be classified into 3 categories. Here is the plan to add these examples back once they are compatible.

Category 1: Using Old Gluon APIs

Influenced examples:

• example/gluon/audio
• example/gluon/lipnet
• example/gluon/lstm_crf
• example/gluon/sn_gan
• example/gluon/tree_lstm
• example/restricted-boltzmann-machine

The old gluon APIs, like hybrid_forward, self.name_scope() are deprecated during the MXNet2.0 developement and most of the changes are summarized in the migration guide. Adapting these examples is straight-forward, just replace the old APIs with new ones.

Category 2: NumPy/NumPy-extension Compatible Operators

Influenced examples:

• example/autoencoder

Some legacy operators are still registered in mx.nd namespace and autoencoder will use nd.UpSampling to build the network. To fix this, we need to register UpSampling in npx namespace.

Category 3: Pretrained Models with old Symbol

Influenced examples:

• example/automatic-mixed-precision
• example/gluon/dc_gan
• example/gluon/embedding_learning
• example/gluon/style_transfer

These examples are using the get_model() API provided by gluon.model_zoo.vision. If the pretrained option is on, it will download the pretrained weights, of which the parameters' names do not match the current parameters' names because of the changes in naming mechanism. We may need to update the pretrained weights.

Others:

example/gluon/word_language_model is using deprecated API: gluon.contrib.data.text
example/multi_threaded_inference: Moved to cpp-package/example