I wrote this as a regression test for the bug fixed in #8807, but seems
like a useful test in general.
Without the fix from PR #8807, this readily fails with an assertion
failure or other errors.
XXX: This is failing on 'main', even with the fix from #8807, just not
as quickly. This uses a small neon.relsize_hash_size=100 setting; when
I bump it up to 1000 or more, it doesn't fail anymore. But this
suggests that it's still possible to "overwhelm" the relsize cache
