docs(fdfd): parallel ILU(0) measured on our matrices, both halves, and not used

[tachsin]

Closes #141, by measurement: neither half of a deterministic parallel ILU(0) pays on photonoxide's matrices, so neither is used.

The factorization (Chow & Patel 2015, doi:10.1137/140968896)

Not built, measured first (comment on #141): the multigrid's smoothers already factorize slab by slab in parallel (about 1.3 s of Diel's hierarchy), and QMR + ILU(0) on the guide spends 4.0 of 5.6 s in its iterations. Chow and Patel's own results (their Section 4.2, Table 3) need 3 to 5 synchronous sweeps on matrices far from diagonally dominant.

The triangular solves (Anzt, Chow & Dongarra 2015, doi:10.1007/978-3-662-48096-0_50)

Built and measured: each triangular solve as k synchronous Jacobi sweeps, all rows in parallel, the same bits on any number of threads; on the transposed factors exactly the transpose of the same series (D⁻¹(I − TᵀD⁻¹)ʲ = (I − D⁻¹Tᵀ)ʲD⁻¹), so QMR stays consistent. The 40³ guide (Shin and Fan's operator, stretched PMLs), QMR + ILU(0) to 1e-8, 20 threads:

triangular solves	exact	1 sweep	2	3	4	6	8
QMR iterations	195	1 163	747	730	929	1 782	2 322
time	3.3 s	12.9 s	10.4 s	11.5 s	15.5 s	32.7 s	48.1 s

3.5 times slower at best, and worse with more sweeps: the factors of an indefinite matrix with PMLs are far from diagonally dominant, and the sweeps' non-normal iteration grows before it converges, as the paper warns for such matrices.

What's in this PR

• docs/methods/fdfd-3d.md, "What didn't pay": both halves, the table, both papers in its front matter; ROADMAP resolves the item as measured and not used.
• The sweeps stay in the code, off unless a test turns them on (Ilu0::with_sweeps and IterativeSolver3d::with_ilu_sweeps are #[cfg(test)]), so the measurement can be repeated: SWEEP_CASE=guide cargo test --release fdfd::three::tests::ilu_sweeps -- --ignored --nocapture.
• A test: the swept preconditioner is exactly its own transpose (uᵀ(M⁻¹v) = (M⁻ᵀu)ᵀv to 1e-12 for 1, 2 and 5 sweeps), and n sweeps on a factor of n rows give the exact solve.
• cargo test --release (360 passed), clippy on both crates.