The difference is because we previously don't fit the intercept in Spark 1.3. Here, we change the input `String` so that the probability of instance 6 can be classified as `1.0` without any ambiguity.

with lambda = 0.001 in current LOR implementation, the prediction is
```
(4, spark i j k) --> prob=[0.1596407738787411,0.8403592261212589], prediction=1.0
(5, l m n) --> prob=[0.8378325685476612,0.16216743145233883], prediction=0.0
(6, spark hadoop spark) --> prob=[0.0692663313297627,0.9307336686702373], prediction=1.0
(7, apache hadoop) --> prob=[0.9821575333444208,0.01784246665557917], prediction=0.0
```
and the training accuracy is
```
(0, a b c d e spark) --> prob=[0.0021342419881406746,0.9978657580118594], prediction=1.0
(1, b d) --> prob=[0.9959176174854043,0.004082382514595685], prediction=0.0
(2, spark f g h) --> prob=[0.0014541569986711233,0.9985458430013289], prediction=1.0
(3, hadoop mapreduce) --> prob=[0.9982978367343561,0.0017021632656438518], prediction=0.0
```

Author: DB Tsai <dbt@netflix.com>

Closes #6109 from dbtsai/lor-example and squashes the following commits:

ac63ce4 [DB Tsai] first commit