Medaka

SNP calling, model, and bugfix release release

Fixed
- Workaround short-contig/no-coverage corner case during pileup.
Added
- Prototype SNP calling and phasing, [benchmarks](https://nanoporetech.github.io/medaka/snp.html)
- Add model for improved Flip-flop model in Guppy 2.3.5
Changed
- Rename models to be more logical
- Update to htslib version 1.9 for long cigars

Bug fix release

Fixed
- Fix bug leading to dropping of pileup chunks during loading

Development and performance release

Changed
- Refactor batch queuing in preparation to using keras Sequence
- Asynchrounous feature loading during inference
- Pin version of h5py to work around intermittent errors in saving models

Training and bug-fix release.

Fixed
- Resolve issue with contained chunks during stitching
- Resolve hanging at the end of training
Added
- Improved storage and retrieval of features for better IO
- Training speed improved >10X
Changed
- Switch to CuDNN for GRU layers
- Check presence of `minimap2` and `samtools`
- Provide more feedback on error

Model release.

Added
- Add support for R9.4.1 flip-flop basecaller

Development release.

Added
- Adds build infrastructure for source distributions and manylinux wheels.