Use combined embeddings instead of random nn.embeddings, tokenizer and vocubulary builder.

[x] Add proper inference method without batch_size and with a response as text not as a token.
[x] Add get batch embeddings map function in embeddings manager
[x] Add combined embeddings map functions in embeddings manager
[x] Implement the batch builder class
[x] Implement vocab builder(Token with Category map) with object equality.
[x] Implement a basic response parser(Only print both token and category_map)
[x] Use combined embeddings in the encoder and fix the tests.
[x] Use combined embeddings in the decoder and fix the tests.
[x] Use combined embeddings in the transformer and fix the tests.
[x] Use combined embeddings, batch builder, vocab builder, and response parser in training and fix the tests.
[x] Use combined embeddings, batch builder, vocab builder, and response parser in inference and fix the tests.
[x] Remove trasformer_utils, Tokenizer, vocabulary builder, and nn.embeddings from model architecture.
[x] Refactor ALiBiBI embeddings calculation in the MHA using the embeddings manager. NOT IMPORTANT RIGHT NOW

nerdslab-club / cl_model