v0.1-r7

1. Increased var_pct_full in ONT mode from 0.3 to 0.7. Indel F1-score increased ~0.2%, but took ~30 minutes longer to finish calling a ~50x ONT dataset.
2. Expand fall through to next most likely variant if network prediction has insufficient read coverage (#53 commit 09a7d18, contributor @ftostevin-ont), accuracy improved on complex Indels.
3. Streamized pileup and full-alignment training workflows. Reduce diskspace demand in model training (#55 commit 09a7d18, contributor @ftostevin-ont).
4. Added mini_epochs option in Train.py, performance slightly improved in training a model for ONT Q20 data using mini-epochs(#60, contributor @ftostevin-ont).
5. Massively reduced disk space demand when outputting GVCF. Now compressing GVCF intermediate files with lz4, five times smaller with little speed penalty.
6. Added --remove_intermediate_dirto remove intermediate files as soon as no longer needed (#48).
7. Renamed ONT pre-trained models with Medaka's naming convention.
8. Fixed training data spilling over to validation data (#57).