Image preprocessing

The approach described in this section, image preprocessing into batches of files, relies on the ND4J FileBatch class (https://static.javadoc.io/org.nd4j/nd4j-common/1.0.0-beta3/org/nd4j/api/loader/FileBatch.html), which is available starting from the 1.0.0-beta3 release of that library. This class can store the raw content of multiple files in byte arrays (one per file), including their original paths. A FileBatch object can be stored to disk in ZIP format. This can reduce the number of disk reads that are required (because of fewer files) and network transfers when reading from remote storage (because of the ZIP compression). Typically, the original image files that are used to train a CNN make use of an efficient (in terms of space and network) compression format (such as JPEG or PNG). But when it comes to a cluster, there is the need to minimize disk reads...