VectorAffineFunction easy constructor

[matbesancon]

VectorAffineFunction has always been a massive pain to construct.
This added constructor builds one from a vector of scalar affine functions, which is the mental model most people will have from a VAF.

This can drastically simplify a lot of code

[odow]

I'm in favor of this. It hurts me every time I try to create a VAF.

Thoughts on VectorQuadraticFunction as well?

[matbesancon]

Thoughts on VectorQuadraticFunction as well?

I'll take a shot in the next days, it will be in a separate PR

[odow]

I can do it

[blegat]

I'm not sure we need this since we already have MOI.Utilities.vectorize

[matbesancon]

the constructor is easier to discover and search for, it has zero cost, gives you a way to construct the VAF that is close to the way it's printed, and the mental model makes sense -> constructing a vector affine expression takes a vector of scalar affine expressions.

MOI.Utilities.vectorize or MOI.Utilities.operate are both awkward to use constructors in disguise IMO

[blegat]

We could consider having two names for the same thing but we shouldn't duplicate the implementation. One of the two should call the other one

[odow]

We could document vectorize in the default constructor?

But case in point: I forget about vectorize

[blegat]

Yes, and vectorize should call the constructor instead of duplicating the code

[joaquimg]

I like this PR, I almost wrote it twice recently.
Can we just use vectorize inside this new function?

[odow]

Nah we should go the other way around: vectorize use this constructor. I can update the PR.

[blegat]

Now the code is a bit weird. vectorize was sharing code between creating VQF and VAF. I would expect the code to also be faster as it preallocate the right dimension. I'd move that code to src/function for both the VQF and VAF constructor and then we can keep vectorize in MOI.Utilities that just calls the constructor.

[joaquimg]

A more naive constructor is not terrible for tests as functions would be small. But vectorize can be used in very large expressions in production code, so it should be preallocating.

[odow]

Closing in favor of #2636