WIP: A Prototype for Enhancing the Performance of Data loader

[mtmd]

Original Issue: #630

Summary

This pull request improves the performance of flashlight by offering a new data structure: Sample. Sample transfers the data that it contains to the GPU memory in an asynchronous fashion. The transformations are performed after the prefetch step to ensure a single CPU thread deals with GPU interactions, hence, reducing the context switching and lock acquisition overhead.
This is a WIP. Please see #631 (comment)

Test Plan (required)

bin/imgclass/fl_img_imagenet_resnet34 \
--data_dir=/path/to/your/ImageNet/folder \
--distributed_enable=false \
--exp_checkpoint_path=/tmp/test \
--logtostderr \
--fl_amp_use_mixed_precision=false \
--data_batch_size=128