LARS fixes: intercept, normalization, bugfix

[rcurtin]

This is the end of a very deep rabbit hole. I thought that #3442 was a result of LARS not supporting fitting an intercept and being able to normalize data to unit variance. That's actually not the case, but I only found that out after writing the support, so here it is.

Deep inside the changes is a minor bugfix that takes a little while to explain:

LARS is an iterative algorithm that adds a feature to the model in each iteration. It selects which feature is the best to add based on residuals, and then selects a step size for how much of that feature to add (so to speak). The paper's rule for choosing the step size, in Equation 2.13, forces a positive (specifically nonzero) step size. The step size is chosen such that we step exactly to the point where the next feature should be added.

However, it can sometimes happen that two different features have the exact same residual value, and could equally be chosen as the next feature to add to the model. In this case, the step size should technically be 0: we should add both features at once. I say "technically" because the paper does not actually clarify this, so my claim there is somewhat rooted in opinion, and I think it's possible to disagree on reasonable grounds.

Anyway, the fix implied by my opinion is to relax Equation 2.13 to allow a zero step size, and thus what effectively happens is that the model can add two features at once.

This fixes the newly-added KKT checks from @gscarella, which I adapted to add to the LARS test suite.

About the intercept and normalization support: I enable them by default since the paper's theory assumes that. There's a little bit of extra accounting to make sure predictions are still right when this happens.

[rcurtin]

This is the tiny relaxation that is the bugfix part of this PR.

[rcurtin]

It turns out the "tiny relaxation" is not so tiny: on LARS iterations where we are using the LASSO modification, sometimes we remove a feature because its sign changes (this is Eq. 3.6). However, we need to avoid considering it for the next iteration---so we can't use the relaxation that allows gamma (the step size) to become 0, since the removed feature's correlation should be exactly zero.

[rcurtin]

I also had one more thing to fix here: LocalCoordinateCoding and SparseCoding are written to pass a precomputed Gram matrix that assumes no normalization and no intercept, which interferes with the changes to behavior in this PR. So, I fixed that, then also added some checks that throw an exception if the user gave a Gram matrix that does not match the settings of fitIntercept and normalizeData. Wish I could have my day back, but it was an interesting adventure nonetheless.

[mlpack-bot]

Second approval provided automatically after 24 hours. 👍