Why Zero-Knowledge protocols are used for NP problems if IP is the class of interactive proof systems where they come from?

The reason is that essentially, the class of languages in $\mathcal IP$ that are not in $\mathcal NP$ cannot be proven with an efficient prover. Since we are typically interested in the cryptographic context with efficient provers, the study of ZK typically focuses on $\mathcal NP$. (Note that when studying complexity and often in foundations of cryptography, we are certainly interested in inefficient provers.)

In order to see why outside of $\mathcal NP$ the prover cannot be efficient, note that any interactive proof with an efficient prover (given a witness) can be emulated by a machine who receives the witness and runs the proof locally, testing if the result is accept or reject. This is exactly the class of languages $\mathcal MA$. Under standard derandomization assumptions, $\mathcal MA = NP$. Thus, under these assumptions, any language that is not in $\mathcal NP$ can only be run with an inefficient prover (even giving it some witness). As such, it is of less interest when constructing protocols and the like.