Decouple Avro specific functions/classes from the rest of Photon-ML about photon-ml

[timyitong]

This PR is one step towards this issue: #126
as we are easing the input format assumption for the simple linear model driver .

[fastier-li]

@XianXing - is this still an issue we need to look at? Please let me know - we might work on it soon.

[XianXing]

I think recent refactoring (client/core/lib) on Photon should be related to the idea proposed here, basically we'd like to make sure Photon is not restricted to Avro serialized/formatted data.

But you guys have more context than me now, so please feel free to close this issue once the client/core/lib refactoring is done (is it done already?).

[fastier-li]

Yes, the refactoring is done. We'll talk about internally about other formats than Avro before I close this ticket, and I'll post the outcome here.

[fastier-li]

The "master plan" is to provide a library of data readers for various format that would all feed GameEstimator.fit with DataFrames and ancillary data structures. This is already exemplified by [AvroDataReader](https://github.com/linkedin/photon-ml/blob/8702fea9a436d4979a3e672fe20ae30e32078f43/photon-client/src/main/scala/com/linkedin/photon/ml/data/AvroDataReader.scala).

I'm going improve on the recent client refactoring by isolating AvroDataReader in an avro package, and colocate in that package the model writer too. We can then have parallel modules for parquet ... and others.

[ashelkovnykov]

While we don't have support for Parquet or LIBSVM data in the client code, sufficient decoupling has been done. For a while now, the input to the GameEstimator has been a DataFrame with features stored as Vector objects.

For that reason, I am going to close this issue and separate ones can be opened for adding specific kinds of readers derived from DataReader.