MultiQC can run out of /tmp space on execution nodes on shark, leading
to incorrect results or crashes. This commit fixes this problem by
specifying a default location for the python tempfiles that MultiQC uses
via the environment variable TMPFILE inside of the shell block for the
MultiQC rule.

This is supposed to speed up MarkDuplicates by up to 16%, see
https://github.com/broadinstitute/picard/issues/902 for details.

This reverts commit 0cf92f4b.

bwa-mem2 can crash when multiple processes are running on the same
system. See https://github.com/bwa-mem2/bwa-mem2/issues/88 for details.

- Start every entry on a new indented line
- Make the order of the entries consistent in every rule
- Ensure equal signs are surrounded by spaces

Add test to make sure the markdup and baserecal groups receive the
correct inputs when a sample has multiple readgroups.

The directory output for the fastqc tasks is causing issues on the
shared file system of the cluster, since it cannot properly determine
the age of the folder. As a result, it re-runs the fastqc tasks every
time a workflow is restarted, regardless of whether the task has already
completed.

To prevent this, a single dummy output file '.done' has been added to
the fastqc tasks which will be written when fastqc exits successfully.

The base recalibration step of the pipeline can take up to 7 hours for
WGS samples, which is a significant part of the total run time.

The developers of GATK state that BQSR requires at least 100M bases per
read group: "We usually expect to see more than 100M bases per read
group; as a rule of thumb, larger numbers will work better."

A human WGS sample with an average read depth of 43x has almost 1300
times that amount of bases. The analysis of these samples would be sped
up greatly by restricting BQSR to a single chromosome.

Use 8 cores instead of just 1, according to the documentation speed
should almost scale linear with the cores provided.

Also reduce the compression level on the output file, since most time
for cutadapt is spent re-compressing the data after trimming the
adapters.

This reverts commit 1b7d807f

This should no longer be needed on the slurm cluster, where each task
can request the amount of tmp space it requires explicitly.

Removing the tmp from the shared filesystem back onto the host running
the analysis should also improve the performance of the pipeline.