Enhance the response classification on dataset based tests

[guy-ps]

Today the dataset based tests rely on the response classifier not contains_refusal_keyword() but this classifier is very basic for some of the advanced needs of the next few datasets that will be implemented.

What is going to be done is:

• We'll enhance the rule based mechanism that checks responses
• We'll implement a semantic mechanism to go over the response
• We'll allow for a choice between the two in the attack configuration