Extra NaN check for conjugate gradient ALS model

With synthetic data and a large regularization parameter, the CG ALS model
would converge so that some users/items factors had 0 vectors for solutions.
The CG update would fail in this case setting all the factors to NaN.
(#106).

Fix by detecting when this would occur and aborting. A previous
check handled the case inside the loop:
#19 (comment)