Fix sequence clipping bug in tokenizer

[justin-barton]

Currently when using the tokenizer's batch_encode method with tensors and special tokens returned, the last two residues are truncated from the longest sequences. For example:

from protein_lm.tokenizer import AptTokenizer

tokenizer = AptTokenizer()

sequences = ["LAGERT", "SERPK"]

batch_encoded = tokenizer.batch_encode(
        sequences, 
        add_special_tokens=True, 
        return_tensors=True
    )

print(tokenizer.decode(batch_encoded[0]))

Outputs:

<cls>LAGE<eos>

[justin-barton]

/take