Performance of GramMatrix -> AlgebraElement

If we use `SA.AlgebraElement` with sparse coefficients, it's not going to exploit the mutability of the underlying MOI expression so it will be inefficient.
The advantage of https://github.com/jump-dev/SumOfSquares.jl/pull/355 where we first determine a the basis with `_NonZero` as coefficient (in which case sparse coefficients is very appropriate) is that we now have a basis where we actually want a dense vector of coefficients.
When the basis is fixed like the sampling basis, it's the same.
So in the `Variable.KernelBridge`, we should use a dense vector of coefficients and make sure `operate!!` is called with `add_mul` on them.

Start in Variable.KernelBridge and trace the GramMatrix to AlgebraElement path, focusing on how coefficients are stored and how operate!! is invoked. Compare the fixed and _NonZero basis cases; done means the kernel bridge uses dense coefficients and calls operate!! with add_mul without changing the resulting expressions.

Written by the indexing model from the issue text.