Fast tokenizer bad merging

[Theodotus1243]

As I see here The default tokenizer behaviour is set to use_fast

But in case of openchat/openchat-3.5-0106-gemma (maybe all gemma models) the initialization of tokenizer goes wrong and <|bos|> token index moves 1->255999, example here tokenizer_config.json

It is a problem of openchat model, but fast tokenizer for Llama always was problematic

So maybe it's reasonable to add flag in mergekit use_fast = true/false?