Clarify policy around which demographics to exclude in Rekognition IP

[AetherUnbound]

Description

Although the final Rekognition IP (#4417) was merged, @stacimc pointed out that we still need to add clarity around the demographics we'll be filtering: #4417 (comment)

Sara and I both singled out the religious affiliation in particular as being problematic for that reason (presuppoing something that can't be determined just from looks), and Zack's and my last comments are both in support of filtering all demographic information. I agree that these are super abstract categories and that it should be considered case-by-case in the review of the tags, but won't the reviewers be using this IP to make their judgments? Explicitly singling these categories out as okay/not worth filtering in the IP does feel like a strict policy as opposed to leaving it to case-by-case review, and if I was reviewing the list of tags I would assume this meant we've agreed to definitely include religious affiliations, for instance 🤷‍♀️

If it's just the case that Rekognition doesn't have any tags related to these categories anyway, so it's not a problem this time, I would still want our documentation to accurately reflect the sorts of tags we want to filter when we eventually encounter them.

We should likely clarify that any demographics should be accepted on a case-by-case basis (with the default being exclusion) based on the set of labels available for a given machine labeler.