PERF: evalbase for spline could be much more efficient

[sglyon]

There are a number of things that could be sped up here. Most of them revolve around allocating less

• We could cache augbreaks based on minorder
• We can avoid calling vec so many times in the call to sparse
• We could try to not even use the sparse method and directly call the SparseMatrixCSC constructor
• Change things like collect(1:m)'' to reshape(1:m, 1, m)
• Don't make B a cell we know what it will really be (Vector{SparseMatrixCSC{Float64,Int64}})
• Roll out loops instead of any [:, ind] = stuff[:, ind]
• Use things like @inbounds or @simd where appropriate

I'm sure there are more, but that is what I came up with from a first look at the method through performance lenses