Refactoring into submodules

[blegat]

The MOI wrapper wasn't yet aligned with the fact that we now have bridges in top of the DiffOpt models.

Closed #220

[matbesancon]

Why introduce submodules? Overall I find it to quickly complicate both internal and external development, I know this is common in MOI but DiffOpt is small enough to contain everything

[blegat]

We have this structure where for QP and CP, we have a Form, Model, Cache, ...
We need to give them different names to avoid clashes so we are adding references to Quadratic and Conic to differentiate them but it's a bit messy. The modules provide a namespace which makes it a lot cleaner.
Second, the DiffOpt optimizer is modular in the sense that it could work with any DiffOpt.AbstractModel subtypes that show how to differentiate for a particular model. We advertise it in the paper and the fact we can refactor cleanly with modules is a consequence of it.

Could you elaborate on why it makes things complicated ? I agree that it makes it complicated if these things are not independent but in this case, they are quite independent. We want to have a clean interface that someone can implement to add new DiffOpt.AbstractModel. Having these module makes this interface apparent and forces us to keep a clean one.

[codecov]

Codecov Report

Merging #219 (2502b59) into master (7094aed) will decrease coverage by 5.41%.
The diff coverage is 66.18%.

@@            Coverage Diff             @@
##           master     #219      +/-   ##
==========================================
- Coverage   91.21%   85.79%   -5.42%     
==========================================
  Files           8       11       +3     
  Lines         842      880      +38     
==========================================
- Hits          768      755      -13     
- Misses         74      125      +51     

Impacted Files	Coverage Δ
src/bridges_unused.jl	0.00% <0.00%> (ø)
src/objective_container.jl	0.00% <0.00%> (ø)
src/moi_wrapper.jl	84.46% <72.22%> (-0.94%)	⬇️
src/ConicProgram/ConicProgram.jl	94.11% <92.30%> (ø)
src/DiffOpt.jl	100.00% <100.00%> (ø)
src/QuadraticProgram/QuadraticProgram.jl	97.12% <100.00%> (ø)
src/diff_opt.jl	100.00% <100.00%> (+0.78%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7094aed...2502b59. Read the comment docs.

[matbesancon]

Major hassle with submodules is the import ..Foo or import .Bar but I see the point that it's not too relevant here

[blegat]

Yes, since the inclusion in the modules shorten the names, the name of the modules + name of the object does not increase too much in length

[joaquimg]

we might need to check both optimizer and the underlying diff_models

[matbesancon]

ping @blegat ?

[blegat]

I'd rather add special cases for the attributes that are used with the DiffOpt.AbstractModel, that's what I just did in the last commit

[matbesancon]

@blegat can you check the two comments above, then good to go

[matbesancon]

bump

[matbesancon]

@blegat good to go I think? We should tag a breaking release right after