Multi-threading bugfix

[cortner]

forces! appears to produces incorrect results when used with ACE1.jl - only for pair potentials, not for many-body potentials. All quite weird. This PR replaces the lock with a numthreads x collection of forces, which might be preferrable anyhow. Not very elegant yet, but works and is needed to make ACE1.jl compatible with latest JuLIP.jl

@tjjarvinen -- I'd be grateful if you can take a look at this.

I'm tempted to just

• merge this
• tag a new patch release
• so I can then next tag a new minor release for ACE1.jl.
• and file an issue that this needs to be looked at again and cleaned up; but this can be done in the process of cleaning up multi-threading in general which we wanted to do anyhow.

[cortner]

@tjjarvinen -- Given I've asked you to take charge of fixing the multi-threading in JuLIP going forward, I don't want to merge this without your thoughts. I know it is a crappy fix, the question is whether you are happy to have this fix in there for now. Should we have a zoom call this week to discuss this?

[tjjarvinen]

The operation seems to be a reduction, maybe using threaded sum operator from Folds.jl might be a better option.

I would also unite the num_threads == 1 and the multithreaded cases. Like, if the multithreading is robust then it should work well with what ever number threads there are, including one.

[cortner]

Main question is still whether to merge this now in order to move forward with ACE1, or work on a more elegant and general fix for multi-threading.

• Re Folds.jl : if you are confident this is a good route to take, then please do try it out, maybe a separate PR?
• Re the numthreads == 1 special case : I know it shouldn't be there. There were some cases a while ago where threading was not robust, hence we have this distinction. This was early days of threading in Julia, JuLIP and ACE, so it may well be fine to remove it now, I just never had the time to test this carefully. I propose to leave this until you rework the multi-threading in JuLIP.

[tjjarvinen]

If it fixes the bug, then lets go ahead with it and implement more elegant version later.