This is a complex topic which was discussed endlessly in the early days of Crossref (and may well still be being discussed). The key thing is to be clear about the semantics of the object you are referencing with the DOI, and the purpose for which you are referencing it.
Also, not all DOIs are equal, and the different DOI Registration Agencies (RAs) have different policies and views on this. I think where Crossref ended up is that if one version of the content is different enough from another than it would matter in a citation which one was being referenced, then it should have a different DOI. In general, it is accepted that preprints are likely to have material differences from final published articles, but differences between AAMs and VoRs are more likely to be related to formatting rather than meaning. However, this likely depends on the policy of any given publisher.
This old blog post on the topic is still interesting to read,
One thing I think it is generally agreed upon - it’s not a good idea to assign multiple PIDs to the same object just because it happens to reside in multiple locations - is this just creates complexity with both referencing and finding an object. There are better mechanisms to deal with this, such as DOI Multiple Resolution. This would imply the need to coordinate among repositories if several hosted copies of the same AAM.
I’d be interested to hear from any of the DOI RAs on their current thinking on this topic.