Make clean docker image routine more efficient

More efficient in regards to our current way of building some specific images which are
growing forever (up until we got no more disk space on thyssen).
That's a simpler workaround than what's discussed in issue [1] [2]. We keep the last "2"
days images so the next build still benefits from it but we drop the rest (so no more
too much disk cost). As this job is triggered daily, that should keep those images in
check.

Provided, the following script is installed on the node:

```
$ root@thyssen:~# cat /usr/local/bin/clean-docker-images.sh

set -x

today=$(date --date '13:00' +%Y%m%d)
yesterday=$(date --date 'yesterday 13:00' +%Y%m%d)

images_to_drop=$(docker image ls \
    | grep -E "softwareheritage/(base|web|replayer)" \
    | grep -v $today \
    | grep -v $yesterday \
    | grep -v "latest")

if [ ! -z "$images_to_drop" ]; then
    echo $images_to_drop \
    | awk '{print $1":"$2}' \
    | xargs docker rmi
fi

docker system prune --filter 'label!=keep' --volumes --force
```

Refs. swh/infra/sysadm-environment#4846

[1] swh/infra/sysadm-environment#4848

[2] swh/infra/sysadm-environment#4846

jobs/tools/docker-clean-images.yaml

@@ -11,4 +11,4 @@
       - timed: "@midnight"
 
     builders:
-      - shell: docker system prune --filter 'label!=keep' --volumes --force
+      - shell: /usr/local/bin/clean-docker-images.sh