Replace Embedding to use nn.Embedding from pytorch

[ravinkohli]

Due to our current implementation of Embedding module, we are forced to one-hot encode all categorical columns. This leads to an explosion in the memory usage. We can avoid this by using nn.Embedding from pytorch which has the same functionality but does not need one-hot encoded columns. This will also allow us to one hot encode the columns which have number of categories less than min_categories_for_embedding.