Also introduced minor modification recently made for other pipelines.

Currently only the folds computed by DESeq2 are available for RNA-seq.

Made paired scatterplot plotting function part of libhts as
`plot_paired_scatters`. Modifications in small RNA-seq pipeline not
tested.

The genome is now configurable. This lead to large modifications in the
config file.

Output organization simplified.

Factorized linking to raw data.

It turns out that using `include:` can fail when there are symlinks in
the path to the snakefile.

This makes the DE results easier to compare with those from small
RNA-seq.

For RNA-seq and GRO-seq, log and data directories are now subdirectories
of output directories. local_annot_dir also moves as subdirectory of
GRO-seq output.

This seems to happen on data matrices with only 2 rows left after
filtering.

The goal is to be able to factorize the rule in a wrapper.

Started to apply to RNA-seq and GRO-seq technique used in small RNA-seq.

I should check to be sure that this is the correct thing to do. The
other possibility would be to sum using agg:
https://stackoverflow.com/a/35403903/1878788

lfcMLE may bring problems with low-expressed genes.

Also:
- Boxplots with labels turned 90 degrees
- do_deseq2 returns size factors

It currently fails when libraries have too many rRNA reads:
https://github.com/arq5x/bedtools2/issues/565

Also added lfc distribution and improved MA plot.

Also moved save_plot to libworkflows.

The older family-level counts are now named "_families".

The column was only added later, when grouping the results with the
counts.

Library standedness was not taken into account correctly, and too many
reads in repeated regions were discarded in featureCounts:
<https://bioinformatics.stackexchange.com/q/2101/292>

Also added mean across replicates for bigwig files and use various log
fold change thresholds to classify genes.

The normalization for the bigwig files is currently based on
reads forward-mapping on protein-coding genes.

Bigwig file generation uses a new method implemented in bam2bigwig.sh
which is faster than bamCoverage from deeptools. It generates slightly
different results (see c5bbd282) and
has not been tested for compatibility with further deeptools-based
analyses..

Also fixed missing pandas import in libworkflows.