Improve turbomind's prefix cache

[lvhan028]

This Pull Request implements several improvements to the tm engine:

• cache prompt tokens and generated tokens when prefix caching is enabled
• remove stateful inference
• update the chat API and CLI, get_logits and get_ppl accordingly
• enable_prefix_caching defaults to True
• merge pt's chat.py and tm's chat.py into one

Remaining refactoring will be processed in another PR:

• remove get_prompt in model.py
• deprecate the built-in chat template, and AutoTokenizer's apply_chat_template instead

test cases:

• llm model evaluation
  https://aicarrier.feishu.cn/wiki/JxoDwiOO7i0xnvkZHzxc3PdGnJh?sheet=95YWnW