optimizing ftrl when calculating w at the last step.

[leleyu]

We use two PSF (FTRLWUpdater and Compress) to calculate w. These two PSF allocate some intermediate arrays, which is memory-consuming when the dimension of model is high.
We can merge these two PSFs into a single one, which directly filter the non-zero ones of w from z and n and save non-zero items into a new vector.