[FEATURE] Basic guardrails for remote model input and output

[ylwu-amzn]

Is your feature request related to a problem?
LLM may generate biased/impolite/hate outputs. We need to build guardrail to block such result.

What solution would you like?
Add basic guardrail based on stop words and regex. If model input and output contains stop word or regex, will throw exception. This will be released in 2.13.

What alternatives have you considered?
No

Do you have any additional context?
We plan to add guardrail based on other guardrail service and semantic after 2.13.

[chishui]

Can this guardrail be turned on and off by users?

[jngz-es]

#2209