feat(bench): the bandwidth each iterative solve reaches, its bytes counted by the kernels, against the triad

[tachsin]

Closes #142.

What

• src/traffic.rs (crate-private): one counter of the bytes the iterative kernels move, by a stated model: each stored nonzero's value (16 B) and column index (8 B), each row pointer (8 B), each vector read or written (16 B a value), a gathered x counted once. Each kernel adds its bytes once a call, from all threads; the formulas are small functions, tested by hand.
• Counted where the kernels are, so the count stays true when they change: the sparse product, QMR's fused passes (symmetric or not), GMRES's sums and combinations, the triangular solves (exact and swept), the multigrid's transfers and corrections.
• bench::Phase::bytes (serde default, so older JSON still reads) for each iterating phase, Measurement::gigabytes_per_second, bench::bytes_moved.
• photonoxide bench's report: a "Bandwidth (of the triad)" column, the GB/s reached and its share of the triad measured on the same threads (the roof, Williams et al. 2009, doi:10.1145/1498765.1498785), and the model in the report's header.

What it shows (docs/benchmarks.md, Core Ultra 7 265K)

problem	1 thread	20 threads
guide-qmr (symmetric QMR, 40³)	19.4 GB/s (59%)	84.4 GB/s (150%)
guide-ilu	15.3 GB/s (46%)	21.6 GB/s (38%)
guide-multigrid	12.7 GB/s (39%)	21.6 GB/s (38%)
diel-multigrid (864 k)	13.9 GB/s (42%)	37.1 GB/s (66%)
strip-ports-multigrid (1.8 M)	13.3 GB/s (40%)	34.3 GB/s (61%)

• Above 100% on the 40³ guide: the count is the kernels' traffic from memory or cache, and the guide's vectors (3 MB each) fit in the 30 MB L3, where only its matrix (about 60 MB) streams from memory; the triad also counts 24 B an element where write-allocate moves 32. The header says so. The larger problems settle at 61–66%.
• QMR + ILU(0) at 38% on 20 threads: its sequential triangular solves (docs(fdfd): parallel ILU(0) measured on our matrices, both halves, and not used #150 measured that sweeps don't help).
• The plan's estimate of "about a fifth of the bandwidth" for the old QMR is replaced by measurement.

Checked

• traffic's byte formulas by hand; a product and an ILU(0) solve add at least their bytes (other tests' kernels can only add).
• cargo test --release (363 passed), the studio's tests (35), clippy on both crates, rustdoc with -D warnings. docs/benchmarks.md regenerated; ROADMAP's harness item updated (the PARDISO/MUMPS export, feat(bench): matrices exported for PARDISO and MUMPS as external baselines (needs the owner's licence decision) #143, still waits on the licence decision).