Why did non-von Neumann languages never became mainstream?
Probably because non Von Neumann machines never really became mainstream. We already have a certain model of the machine in our heads; it's probably easier to implement a language that works in the same way, and to train programmers in such a language.
Of course, maybe the reason VN architecture is dominant is that there are no good non-VN programming languages... but asking why the architecture is the dominant one probably deserves a different question.
Also, I'm not sure there's a clear distinction between languages that are "Von-Neumann languages" and those that aren't. Are functional languages considered VN ones? They deviate from the classic VN model in some regards, but are similar to it in others. And many functional languages are mainstream.
One reason is lack of a killer language: APL is one of the most well-known non-Von Neumann languages, but it's not exactly easy to pick up for most people.
Does Haskell written in point-free style count as non-von Neumann programming language? If so, that's probably the most wide-spread example of current non-von Neumann languages.