[SPARK-15328][MLLIB][ML] Word2Vec import for original binary format

[wangyum]

What changes were proposed in this pull request?

Add loadGoogleModel() function to import original wor2vec binary format.

How was this patch tested?

mllib.feature.Word2VecSuite and ml.feature.Word2VecSuite

I also tested with real model:

[AmplabJenkins]

Can one of the admins verify this patch?

[insidedctm]

This seems to work fine with small model such as that produced by demo_word.sh in the word2vec code repository however I get problems when trying a large model such as GoogleNews-vectors-negative300.bin.

I can successfully load the model using this code (albeit I needed to give the driver 12GB of memory):
import org.apache.spark.ml.feature.Word2VecModel
val path = "file:///Downloads/GoogleNews-vectors-negative300.bin"
val model = Word2VecModel.loadGoogleModel(path)

However synonyms are not found for a typical lookup e.g.
model.findSynonyms("spark",20).show
responds with
java.lang.IllegalStateException: spark not in vocabulary

However the distance tool from the word2vec toolkit, loading the same model gives: