fix overflow when model size > 2^31

closing file f_in in ffm-predict.cpp

Windows support

restruct

inner iteration should be limited to order.size()

change initial values of latent factors

support auto-stop

compilable on Cygwin

support on-disk learning

compilable on Windows