Bugfix train tokenization qwen

[maxime-louis]

Label delimitation problem fix with Qwen:
Mistral/llama/solar/gemma tokenizers all have:
user_message + generatino_prompt + answer != chat_template(user_message, answer):
As they always add an extra newline in the second case. Qwen doesn't do that, so our way of using the length of "user_message + generatino_prompt" (:= position) to delimitate the label for training does not work.

With this fix:

• we try to recover the label at (position) and (position - 1)

=> Behaviour is unchanged for all models but Qwen
=> Alexandre ran multiple experiments with Qwen and that code: all works well, no warning.

Qwen pad_token handling
Qwen has no bos_token, which we used as padding in bergen: we take pad_token if it exists (it does for Qwen), and eos_token in last resort.