Short answer: Rasa NLU is flexible in a way where one can choose to make them non contingent as per the need.
Long answer: For a very minimal language task, one can choose to not introduce entities and limit models to perform only intent classification. On the other hand, entity classification (or NER in the context of conversational AI) is a relatively hard problem (given small dataset or large unlabelled dataset) and not completely solved.
You can dig deeper here in NLU Pipeline docs to understand how intent and entity classification are not hard-wired and you can easily decouple them.
An answer specific to your example would be to think of a scenario where “how about Newyork” is a question asked after user has asked “whats the weather like in SF”. The assumption would be that bot is able to give a meaningful answer to the first query and user asks a succeeding question. In such case, knowing the state of the conversation and preceding question would make sense.
More importantly, it would be helpful to understand is your objective whether it is conversational AI or just NLU. IMO the answer to your question would vary in Academia Research vs Solving a real world problem. But digging deeper into intent classification would help for sure.
Hope this helps!