Learnmsa

Added language model support `(--use_language_model)` for significantly improved accuracy.

With this option, learnMSA aligns about 6% points more columns correct on average on HomFam benchmark than state-of-the-art tools including learnMSA without LM support.

Added sequence cropping to accomodate very long outlies.
Previously, learnMSA would be very slow, if the input includes sequences many times longer than the average. This release provides a fix by implementing the command line option `--crop x` to crop inputs longer than x (defaulting to 3 times the average). Cropping can be disabled (the original behavior) with `--crop disable`.
Cropping does only affect model training, not the decoded MSA which will always respect the full input.

Fixed missing package data for the `--use_language_model` option.

- improved numerical stability across TF versions
- minor bugfixing

- Use `pyfamsa` to align insertions, also made aligning insertions the default behavior (also added `--unaligned_insertions` flag).
- Use `biopython` for data parsing. Many more input file formats are now available as well as the experimental `indexed_data` flag for large datasets that allows constant memory model training.
- Multi GPU training works now.
- Added the highly experimental `--use_language_model` flag that uses a large, pretrained protein language model to guide the MSA and improve alignment accuracy.