[tannevaled]

Moves ndarray's speed comparison onto go-scikit/refbench v0.1.2 and replaces the multicore page.

What changes

• bench/ (own module, as perf: migrate the SciPy comparison to refbench v0.1.1 (#33 step 1) go-numerics/special#35 and go-simd/floats bench/perf): rbtest.Perf, 256 rows covering Array methods, constructors, fft, linalg, random (Generator, RandomState, package functions), one case per public function (several shapes of use get several rows) and 121 named exemptions with a reason in docs/perf-exempt.md.
• One core on both sides (GOMAXPROCS=1, one BLAS/OpenMP thread in the reference worker), NumPy pinned at refbench's 2.5.4 (bench/refs.txt), interleaved rounds, load gate. fft rows include the plan on both sides.
• Data: every array is built in Go and NumPy from the same expression; TestDataParity compares the SHA-256 of both sides (24 distinct argument sets, complex ones included, all equal) and TestRefSmoke runs every NumPy call once (go test -perf.parity, run here before opening this PR).
• docs/perf-table.md and docs/perf.json are generated. The table is deferred: the host was at a 1-minute load of 8 to 80, so only allocation counts are recorded, no timing verdict. A timing run at a quiet load replaces it.
• docs/perf.md is now the method page; the old multicore page is kept as docs/perf-history.md with a banner (its kernel analysis is still referenced). README.md no longer quotes the multicore ratios; BENCHMARKS.md carries a "not a verdict" banner.
• CI: new perf job (allocation counts, table regeneration check, public-function inventory), with GOMAXPROCS=1 in the environment because go-fft/fft reads it once at init (parallel.go reads GOMAXPROCS once at init: N-D transforms stay multi-core after GOMAXPROCS is lowered go-fft/fft#107).

Issues

• Closes perf: docs/perf.md compares multicore Go with single-threaded NumPy 2.2.4; re-measure on one core against the pinned NumPy #93 (the page that compared multicore Go with single-threaded NumPy 2.2.4 is replaced; the multicore column is not carried over).
• Refs perf: random Choice allocates twice per drawn element (200 013 allocations for 100 000 draws) #103 (Choice allocates twice per drawn element: 200 013 allocations for 100 000 draws), filed from the allocation counts in this PR; not fixed here.
• Refs perf: linalg SVD, Pinv, Lstsq, Eigh, QR, Eig allocate 90-276 times per 64x64 call #104 (linalg SVD/Pinv/Lstsq/Eigh/QR/Eig allocate 90 to 276 times per 64 x 64 call); not fixed here.
• Refs parallel.go reads GOMAXPROCS once at init: N-D transforms stay multi-core after GOMAXPROCS is lowered go-fft/fft#107 (GOMAXPROCS read once at init), the reason for the environment variable.

The timing measurement of the 256 rows (and an issue per slower row from refbench issues) is the follow-up once the load allows it.

🤖 Generated with Claude Code