Clean all logs after recovery

[Tpt]

Allows to make sure no old log is staying after proper recovery

Fixes #145

[Tpt]

I am not sure it's a clean fix of #145. It prevents more a cause of the bug than fixing it really.

[arkpar]

Could you also add a test from #145 please.

The issue is indeed that the old log was not removed and was replayed, screwing the header on the value table. There's a TODO to detect this issue at validation, which should have prevented it as well, once implemented.

When validating, logs are removed out of order. Each successfully validated and enacted log is added to the cleanup queue. But if there's an error, the erroneous log is dropped immediately and the cleanup queue is not processed at all.

While this fix does indeed cleans up everything, it is still in the wrong order. So if an (unlikely) crash happens when the logs are being cleaned, the bug would still be there.

[Tpt]

@arkpar Thank you for your detailed explanation. So, I understand correctly the fix to do is to implement the TODO to detect this issue at validation.

And, maybe, process the cleanup queue in case of a validation error. But if we have the validation then it should not be required anymore if I understand correctly.

I am closing this PR and I'm going to write an other one.

[arkpar]

Fixing the TODO wont be enough in the general case. If you don't change the number of value table entries, but just modify an existing value, you'll end up with a slightly different version of this bug. The log would be valid, but the value will be old. So I would not bother with it for now.

I suppose the most relevant fix would be to

1. Not drop the log in clear_replay_log, but instead add it to the cleanup_queue
2. Process the cleanup queue even the database is shut down with error in kill_logs.

[Tpt]

@arkpar Thank you!