investigate padding issues

[mivanit]

in #195 , a padding issue due to a new transformer_lens version was fixed -- but apparently, left padding was not supported. Makes sense to investigate whether this makes any difference:

• train a models with both old and new (>=1.6.1) transformer_lens version, same config
• check if they work any differently under the same transformer_lens version they were trained on
• see if behavior changes when updating transformer lens

somewhat related to #121