Spectral fixes

[Shannon Quinn]

I have a load of fixes in the pipeline for the spectral clustering 
algorithms. The work on Eigencuts is extensive and still ongoing, so 
while I will post those tickets, the fixes will likely not make it for 0.8.

SpectralKmeans, however, has numerous fixes that are ready to go. Before 
I post and commit them, I would like some input on the following items:

1: We added the option to use SSVD in place of the Lanczos solver. Would 
it be acceptable to have a command-line flag to specify the solver to use?
2: Lots of temporary files are generated by the numerous MR jobs chained 
together. Is there a rule of thumb for whether or not to delete these 
intermediate files after running the whole job? Right now I have a 
command-line flag to indicate whether they should be removed or not.

Thanks!

Shannon