use separate batch for snapshot orphans

[erikgrinaker]

Fixes #261. I tried several approaches, this is the simplest one that passes all current tests.

I found it hard to reason about the interactions between orphaning and pruning, and am concerned that this only adds to the complexity, and thus the potential for harmful interactions. I will write a comprehensive, randomized stress test in a separate PR.

[tac0turtle]

utACK, excited for the randomized tests.

[erikgrinaker]

I realized this has at least two bugs:

• Deleting a previously persisted version will flush queued orphans for non-persisted changes.

• Deleting a previously persisted version will not take into account and adjust orphans queued since the last persisted version.

I'll come up with a fix for the former, but I don't believe the latter can be easily fixed without a pruning redesign.

[erikgrinaker]

Allright, I think this should be good to review again - it's not pretty, but it works. I haven't fixed the second bug above, which requires a pruning redesign, but it's not likely to trigger for most use-cases (e.g. the SDK's usage pattern) - the consequence should only be stale orphans being left behind, I think.

The test failure is because of #269.

[erikgrinaker]

The randomized test suite being built on erik/testageddon is showing some problems with this PR, it's unclear if it's caused by this change or if it's an unrelated problem since the suite panics on master. Will convert this to draft until investigated.

[erikgrinaker]

No longer relevant, pruning was removed in #274.