Add compatibility to read in bigmemory data sets

[dustinvtran]

Since we're doing scalable computation, we should work with scalable I/O, storage, etc. packages. That is, we'd like to be able to run SGD on data sets with memory larger than a computer's RAM.