Fix bos/eos tokens in bpe tokenizer in c++

[guangy10]

Those tokens are hard-coded for llama and doesn't work for other models. They should be set dynamically based on the config of the tokenizer to support HF models.

[guangy10]

Fix in #4149