Add tachyon back as dependency

[dcrankshaw]

PR #27 removed tachyon as a dependency to get binary compatibility between hadoop-client and spark-core. Even though we can't use the Tachyon KV store right now, we can still use Tachyon as a faster version of HDFS to transfer training data and models between Spark and Velox. We should add tachyon back as a storage option once we can figure out class compatibility between Spark, Hadoop, and Tachyon.

[tomerk]

As of pull request #44 this may be as simple as changing the spark data location config to be in the tachyon filesystem