Primary changes:

• StackedTransformer: Adjusted the project_to_hidden_size to be initialized with the output shape of the embedding rather than the embedding size. When a sparse representation is specified for the embedding, the embedding's output shape is actually the vocabulary size instead of the embedding size.
• loss_modules: Add support for specifying class weights using a list.
• image_utils: Add fallback for num_channels=None. When num_channels isn't specified by the config, finalize_preprocessing actually reads in the first image (or a sample of images) to determine the number of channels.
• image_utils: During num_channel inference, use the maximum num_channels instead of the average over all samples.
• test_experiment: Updated image reading to use torchvision.io per offline discussion related to TF to Torch Fix integration test for server #1411, and commented out sequence-output related unit tests.
• embedding_modules: Implemented output/input shape for TokenAndPositionEmbedding and added a unit test.