Flexible TFRecords Data Provider

[damro]

Write a tfrecords data provider that

• uses tf.python_io.tf_record_iterator() method to read data without a session
• can post-process data
• loops through files infinitely by using a list of filename queues, one for each source
• starts to read from a new file when it runs out of entries in the old one.
• that allows to read data from separate source and sync them. (This would help to make tfrecords datasets easily expandable.)

[qbilius]

Since all of that can be also done using multiple small hdf5 files, perhaps we should make this provider non-TFRecords specific?

[damro]

Sounds good. I included it in the list of things to do. See 2. at issue #48