Improve JBOD throughput

Nodes with spinning disks have very limited throughputs during compactions. This is mainly due to https://issues.apache.org/jira/browse/CASSANDRA-6696[CASSANDRA-6696] where each SSTable got "sticked" to one data directory only. In order to increase the performance during compactions, reads and writes should always happen on different directories, and a single data directory (for both reads or writes) should be used by no more than one compactor at time. The current throughput of a single spinning disk can get as low as 20MB/s due to r/w seeks penalties, but a compaction between two different spindles could easily get 100MB/s or more throughput. Since 6696 requires sstables to always be on the same data directory, even a compaction between two different data directory and then a "brutal" copy back to the "appropriate" directory could get an overall throughput of 50MB/s. Another option would be to let SStables "float" between two data directories only, and "switch" between them at every compaction.