Remove parsing perf bottleneck in WordEmbeddingsTransform

[adamsitnik]

I am currently profiling ML.NET to find performance bottlenecks and places where .NET framework could do a better perf job.

I started with two the most time-consuming benchmarks from our public benchmark suite.

After some profiling, it turned out that parsing big text files in WordEmbeddingsTransform is a performance bottleneck.

In the histogram below the red box 2 is parsing.

I optimized the code and then parallelized it.

Before:

Method	Mean
WikiDetox_WordEmbeddings_OVAAveragedPerceptron	286.7 s
WikiDetox_WordEmbeddings_SDCAMC	184.1 s

After:

Method	Mean
WikiDetox_WordEmbeddings_OVAAveragedPerceptron	169.02 s
WikiDetox_WordEmbeddings_SDCAMC	65.32 s

The PR is #1599

/cc @shauheen