Figure out how to store corpora more efficiently

We are storing too much corpora from old experiments and need to delete them to spend reasonable amount on storage.
We should consider better strategies in the future such as:
1. Having a retention period after which data gets deleted.
2. Having the runner only store new corpus files at each snapshot (this can be done either by hashing each file or checking its last modified time
3. Deleting most corpus files at the end of the experiment.

I'm leaning towards 2. 2 will also make implementing a queue based measurer much simpler.