If yes, is there any paper that proves it? Unifying Zero-Knowledge Proofs of Knowledge, by Ueli Maurer, argues that Schnorr's interactive protocol is zero-knowledge. If this is true, using the Fiat-Shamir transform, can we convert Schnorr's protocol into a NIZK proof?
Yes, and in fact, Schnorr's signature scheme was originally described as a non-interactive protocol. I think the confusion around interactivity comes from the fact that the same paper first described a interactive identification scheme, which can be viewed as a specialization of the signature scheme for empty messages.
In both schemes, challenges can be generated either interactively or using hashes. The non-interactive variant can be viewed as an application of the Fiat-Shamir transform, although Schnorr did not describe it as such.
Note that while more general formulations of the Fiat-Shamir transform involve a random oracle assumption, the Schnorr signature scheme in particular has weaker requirements -- see Navel et al. and Chen et al.