Add atomic relaxation

[amandalund]

This adds the ability to simulate the emission of fluorescence photons and Auger electrons when a primary process produces a vacancy in an atomic shell.

• Add a reader to load EADL transition data
• Add host/device storage for atomic relaxation data
• Implement atomic relaxation
• Include relaxation in Livermore photoelectric model

[amandalund]

Also did a quick sanity check comparing the secondary energies in Celeritas vs. Geant4 with the Livermore photoelectric model + atomic relaxation (fluorescence and Auger enabled) for a few different elements and incident energies, e.g.:

[sethrj]

Kickass results @amandalund !

[sethrj]

Excellent work. I'm a little uneasy about how we're combining multiple setup-time decisions (no atomic relaxation, relaxation without auger electrons, full relaxation) into low-level runtime conditionals. It's sort of an optimization, so we don't have to worry about it right now, but I think for situations like this the Model and Interactor classes could be templated based on the behavior (maybe something as simple as an enum defined in the Interface class, which can be used with some traits classes to move around the necessary data), then the setup-time Process::build_models would emit a different Model based on whether the selection.

[sethrj]

shell_id should be part of the constructor -- "distribution"-like functors should take the RNG as their only element.

Looking at the implementation of AtomicRelaxation, we might want to add a new small helper class to emit the distribution, which can also connect the dots for "how many secondaries to allocate" and "how to allocate and return gracefully if there's not enough room". Maybe something like this would go into the interactor body:

Span<Secondary> secondaries;
if (AtomicRelaxationHelper relax_helper{shared, el_id, shell_id})
{
    // Need to do relaxation (define operator bool() as num_secondaries > 0)
    secondaries = relax_helper.allocate(allocate_secondaries_);
    if (secondaries.empty())
    {
        // Failed to allocate
        return Interaction::from_failure();
    }

    // Create the sampler from the allocated secondaries
    AtomicRelaxation sample_relaxation = relax_helper.build_distribution(secondaries);
    // Sampler returns the subspan of created secondaries
    auto num_secondaries = sample_relaxation(rng);
    CELER_EXPECT(num_secondaries > 0);
    // TODO: I need to implement "first" and "last" in Span
    secondaries = secondaries.first(num_secondaries);
}

[sethrj]

For now, I think the model interface needs to be device-only since interact always assumes device-only. If and when we want to support transport on the CPU we'll need to implement another interact that takes host pointers.

[amandalund]

I still have a little work to do here @sethrj (helper class to emit the distribution, possibly address your concerns about the setup-time decisions, which I definitely agree with -- uneasy is the right word 😉), but thanks for your feedback so far!

[sethrj]

Looking good! A couple more comments.

[sethrj]

Not gonna lie, every time I read this variable name I think of Karl Urban's character in Thor: Ragnarok 😂
https://getyarn.io/yarn-clip/54bf355e-34fd-4d10-8e82-7a3564ebb1c7

[sethrj]

Still true a year and a half later 🤣

[sethrj]

Nice work @amandalund !

[sethrj]

@amandalund Ready for merge, right?

[amandalund]

If you say so @sethrj! Thanks as always for the helpful feedback.