Take advantage of separable basis functions

[unalmis]

Right now, anytime we compute a series expansion, for all of the basis classes in DESC, the resulting Vandermonde matrix is computed by first forming an outer product of the mode numbers then evaluating the basis functions.

For each point $x$, this generates a large mode matrix $A(x)$ of size $L M N$. If there are $P$ points then the matrix $A$ has size $P L M N$. From this matrix all the basis functions $f \colon A \mapsto f(A)$ are evaluated. In the performance analysis for this type of stuff, the dominant expense is by far evaluating $f$. Hence, this is an inefficient approach to computing the Vandermonde matrix because $f$ must be evaluated at $P L M N$ times.

Now, because the basis function associated with any given spectral coefficient $a_{l,m,n}$ are separable in the spatial dimensions, more efficient methods can be used to generate the Vandermonde matrix. In particular for any basis the $K$ dimensional Vandermonde matrix should be computed by

1. first forming $K$ one-dimensional Vandermonde matrices ${V_i | i \in (1,...,K)}$ in each spatial coordinate
2. Computing the basis functions on these matrices ${f(V_i) | i \in (1,...,K)} $
3. Generate the desired $K$ dimensional Vandermonde by taking the outer product $f(V_1) \otimes f(V_2) \otimes, \dots, \otimes f(V_K)$

This reduces the computational cost from $\mathcal{O}(P L M N)$ to $\mathcal{O}(P [L + M + N])$ and can be done without making any assumptions on the grid or computing auxiliary data. I have already updated the DoubleFourierSeries basis in #1440 with this change and it should be simple to extend it to the other basis.

This can be done in complement to the partial summation techniques discussed in #1154 and #1243 .

[f0uriest]

I think the unique option of Basis.evaluate already sort of does this, and is improved in #1405 to not need repeated calls to np.unique

[unalmis]

1. That option is not JIT compilable or differentiable.
2. Calling np.unique and then indexing the unique modes and expanding them with gather operations are unnecessary and somewhat expensive. It also involves storing more things in memory.