It turns out `batchPromises` and `fallbackPromises` was already doing
the "SQL-only first, relations later" split the mentioned in the
previous PR, so I think this bug was already fixed.

I.e. this PR largely reverts the prior PR, but keeps the reproduction
test case, and it still passes, indicating that either:

a) the repro test case never actually reproduced the bug :thinking:,
b) the repro test case did reproduce the bug, but we've since fixed it

The populateBatchLoader has been heavily rewritten/refactored since
the original bug report, so my assumption is that it's b) and that
prior refactoring happened to fix this bug.