[Rivet] Automated projection ordering tests

[Frank Siegert]

> one thing ahead: I just fixed a nasty projection ordering bug affecting
> Fastjet. All runs that have mixed analyses with/without invisibles in
> the Fastjet clustering are affected. Up to now, the projection applier
> would not take that difference into account, and potentially apply the
> wrong fastjet projection depending on analysis ordering, e.g. using
> visibles instead of invisibles!

Ouch (and good that you found it)! This is a major issue.

I wonder how we can prevent something like this from happening again,
since the projection ordering logic is definitely one of our Achilles'
heels.
Wouldn't it be good to add a test which runs as many analyses as we
can think of in one run, and then the same setup in a second run where
we disable projection caching (I assume that must be fairly trivial to
disable, alternatively we could also run each analysis separately). A
diff of the output histograms should give us a fairly exhaustive
indicator whether the projection ordering is buggy?

I'll reply to the yaml/2.1.0 related stuff in the original thread.

Cheers,
Frank