0.1.8-benchmark: Add ofrecord_image_classification_reader (#3419)

* ofrecord_image_classification_reader

* refine

* use num_decode_threads_per_machine

* Specify image_feature_name and label_feature_name

* Extract LoadWorker and DecodeWorker

* Remove image_feature_name/label_feature_name default value

* underlaying_=>base

* set decode_buffer_size_per_thread to 32

Former-commit-id: 526766a72e7bbb1ebb1897c8dca90c254b1576cf