arrow

ARROW-783: [Java/C++] Fixes for 0-length record batches

@StevenMPhillips @nongli @julienledem I found a number of issues in both C++ and Java around the handling of 0-length vectors. It seems that preserving a single inconsequential offset for a length-0 variable length vector can be a bit difficult, so I relaxed a restruction in `loadFieldVectors` about this. Let me know if there's anything concerning about the other changes around EOS signaling

Author: Wes McKinney <wes.mckinney@twosigma.com>

Closes #505 from wesm/ARROW-783 and squashes the following commits:

28ddcab [Wes McKinney] * Have loadNextBatch return true/false for EOS to accommodate 0-length record batches * Relax n + 1 restruction for 0-length vectors