Updates to Inner Quadrature Grids

[paul-vdb]

@weizhangstats @paciorek I'm updating some of the Laplace/AGHQ methods to go along with the Laplace paper. Here are the basic updates that I think will be useful and have been on my to do for a while.

1. quadTransform_ = "identity" will lead to evaluating the log likelihood at the provided quadrature points only. No transformation. Simple quad sum and simple gradient.
2. quadTransform_ = "center" will lead to taking quadrature nodes and centering them around the RE mode. Minor change to current methods.

I think these two will be useful for things like SCR and other problems where the random effects might not be normally distributed and could be multimodal and the user might just want to use a grid-based quadrature method.

Additional updates:

1. Sparse grid quadrature and user grids being added to AGHQ. This is really just bringing in the changes we made for nestedApprox and allowing them for Laplace/AGHQ marginalization. This is going to include pruning, user supplied grids etc.
2. Wei's gradients for the multivariate AGHQ grid... This one might take me a bit of extra time...

See branch SparseAGHQMLE

[paul-vdb]

Update 1: in 1D quadrature added options 'center' and 'identity' to quadTransform. These options allow for the user to supply the quadrature grid directly and have it used directly in the case of identity and in the case of center, it is centered to the maximum, not scaled. Note adaptive is scaled and centered (cholesky or spectral). Note centering and one node just defaults to laplace because I can't believe there is a single node use case that isn't Laplace...

Testing against using uniform random variables for random effects as a standard use case preparing for an SCR example: