Segfault on load for embedding models lacking BOS/EOS token (such as `BAAI/bge-*`)

For embedding models that lack BOS/EOS tokens (such as `BAAI/bge-*`), the BOS/EOS token ids default to `-1`, which causes a segfault on loading when calling `token_get_text`. I would recommend either short-circuiting these calls to the empty string in that case or skipping the chat template code entirely for embedding models.