Citing papers with newer arXiv versions

I think best practice is to include the arxiv identifier for any paper you cite anyway, so the bibliography will include both the publication info and a link to the arxiv. In that sense you don't have to choose just one. If it's say in the introduction and you're writing "The stabilization hypothesis was introduced in [BD]" there's no need to specify which version you're referring to. However, if you want to refer to a specific theorem number then you might want to clarify if the numbering disagrees in the two sources. Similarly if you want to refer to something that's only in a later revision then you have to be clear. So you could say "Theorem x from [BD] (which appears only in the arxiv version) says..."

In this case it seems like the only change between the 1995 arXiv version and the 2004 one is that some missing figures were added. You can see this in the arXiv comments field (Also note that v2 is still dated 1995). Therefore you should go ahead and cite the published version.

If in the future you find a published paper with unpublished addenda then you can either cite both versions or cite the published version and mention the differences in a footnote. But it's always a good idea to cite the published version so people know it has been peer reviewed. I would find it quite odd to see an arXiv paper from 2004 cited in a paper (pre)published in 2020 because then I'm worried that that paper was never accepted.