Tune the logic used by Magma to hold checkpoints for rollback

Currently magma holds either 10min or 10% of overhead as a fraction of total data size as checkpoints which can be rolled back. The overhead is computed as the amount of extra data held on disk by the checkpoints due to SSTables being deleted and created by compaction. In higher write throughput workloads we have noticed this overhead computation is a bit aggressive and may need to be altered.

We could also investigate kv_engine providing magma with a safe seqno upto which it holds checkpoints. This would avoid rollback to zero completely which can be very costly when the data size is in terabytes.