Investigate apply_chat_template per message vs. all messages

This is an issue to track the investigatory work of #38 .

Before making the repo wide change to use apply_chat_template on all messages vs. per message, we should do some analysis on how frequent the boundary issues are. An example of such analysis would be:

* Take top 25 instruct tokenizers
* Check if there are boundary issues for multiturn (e.g, check `apply_chat_template(messages[1:i]) for i in range(N)`)
* check arbitrary end-strings (ones other than what the instruct checkpoint comes with) assuming users will fine-tune on their new end-string