Add the Parser to parse the data of different schemas to Tensor.

[brightcoder01]

• 1,2,3;4,5,6 => 2 x 3 Matrix

• 1,2:7.0;3,4:8.0 => A rank-2 Sparse Matrix, it only contains two non-zero values: 7.0, 8.0.

• I:1,like:1,apples:1 => A string "I like Apple"

And more schema.

There are two options to implement the Parser:

1. A customized keras layers (provided from elasticdl or user), use native TF OP to implement the parse logic. => We need check the performance.

2. A customized OP implemented with C++. It can be using in the dataset_fn for training and loaded by serving framework for inference.