This PR changes lambda scaling from number of users/items to number of explicit ratings. The latter is the behavior in 1.2. Slight refactor of NormalEquation to make it independent of ALS models. srowen codexiang

Author: Xiangrui Meng <meng@databricks.com>

Closes #5314 from mengxr/SPARK-6642 and squashes the following commits:

dc655a1 [Xiangrui Meng] relax python tests
f410df2 [Xiangrui Meng] use 1.2 scaling and remove addImplicit from NormalEquation