Add libsvm dataset support

[yupbank]

Address #10

add make_libsvm_dataset function, which returns a dataset contains (feature, label) per row.

def make_libsvm_dataset(file_names,
num_features,
dtype=None,
label_dtype=None,
batch_size=1,
compression_type='',
buffer_size=None,
num_parallel_parser_calls=None,
drop_final_batch=False,
prefetch_buffer_size=0):

[yongtang]

Overall looks good, though I am wondering if we could expose a class interface such as class LibSVMDataset(dataset_ops.DatasetSource)? Maybe we could add the class interface on top of the current implementation?

[yupbank]

it is hard, since we only have a parsing kernel for now, we need to implement a datasource kernel to support that basically.

if it is really worth it, i can make a second pr to port current paring kernel into datasource kernel

and the function pattern is also from tensorflow core https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/data/experimental/ops/readers.py#L311

[yongtang]

Overall, I think this is good. We could consider adding DatasetSource support later. LGTM

[yongtang]

@yupbank Added a PR #38 to fix some minor issues. Please take a look.