libFuzzer merge cycle should be more robust against interruptions

[choller]

Currently, a libFuzzer merge cycle with S3 performs the following steps:

1. Delete closed queues
2. Download remaining queues (thereby closing them)
3. Local merge
4. Corpus reupload

If the task gets interrupted in step 3 or 4, the next run will delete the queues closed previously, even though they haven't been merged yet. To make this more robust, we should probably close the queues only after completing step 4 successfully.