Added a feature to VLLM and HuggingFaceLM that forces a response prefix

[junya-takayama]

Add a feature to VLLM and HuggingFaceLM that forces a prefix for the assistant's responses.

Example

from flexeval import VLLM

lm = VLLM(model="/path/to/model", prefix_str_for_chat="<think></think>")
output = lm.generate_chat_response(["role": "user", "content": "hogehoge"])
print(output.text)
# <think></think>fugafuga

[Kotaro-Aono]

問題なさそうです！
LGTM！