[Bugfix] add truncate_prompt_tokens to work offline, directly from LLM class.

[yecohn]

Fixes #4507.

• added function _validate_prompt to make sure prompt is a right format and run some validations.
• truncate_prompt_tokens should take truncate the prompt to the left as seen in vllm/entrypoints/openai/serving_engine.py line 182.

[simon-mo]

@tdoublep can you help review this?

[tdoublep]

@simon-mo sure, will try to get to that later today