This PR is the result of a side-channel discussion with @stedolan - it just documents the current state, which should be fine for 5.0. A separate issue can be opened if wanted to track investigating improvements to it.

If this hash function was a known design described and evaluated in the literature, a comment pointing to the description and evaluation would be useful. A comment that just says "we picked this function on a whim and have no idea how well it works" is not useful, so I'll close this PR. I make a mental note to replace this code by a well-known hash function soon (but not in 5.0).