Ref T7664. Currently, when spreading subpriorities we may recurse deeply in certain conditions. Make sure we never recurse more than one level.
To try to mitigate issues with floating point precision, be more aggressive about selecting tasks to reorder.
I wasn't really able to come up with a realistic test case here, and the test cases I found which sort of approximated the behavior took way too long to generate data to actually commit.
This approach is inherently somewhat fragile but hopefully this is approximately good enough. We don't have a durable storage engine which can meaningfully represent double-linked lists right now.