Support filter push down in rikai.dataset

[eddyxu]

When loading the dataset for training, we'd desire to split the dataset into train, test, eval split. And it should make it easy for an user to just load one of such split, for example

from rikai.pytorch.data import Dataset

train_dataset = Dataset("foo.bar", filters=["split = 'train']) 
eval_dataset = Dataset("foo.bar", filters=["split = 'eval'"])

We could look into the pyarrow's filters in parquet dataset to see whether we can use them.