K1 remainder: float32 kernels, Dot1D and the dot inside GEMM, Maximum/Minimum move to go-simd/floats #98

Tracking issue for the part of K1 of go-scikit/roadmap#27 (docs/performance-campaign.md, section 4.3) that #97 (v0.37.0) did not move into go-simd/floats. Each is a kernel go-ndarray keeps privately; the rule is that a kernel lives once, in the lowest library every user can import.

• float32 elementwise: Bin32 (add32/sub32/mul32/div32: SSE2 on amd64, NEON on arm64, VSX, LSX, RVV, vector facility), Sqrt32 (internal/kernels/bin32.go)
• float32 extrema Max32/Min32 (extrema32*.go), with the signed-zero contract of go-simd/floats (issue filed there)
• AddBlocks8F32, the eight-accumulator pairwise block of the float32 sum (blocks8*.go, used by typed.go's pairwise sum)
• Dot1D / Dot1DP, and the dot inside GEMM/MatVecP (dotFMA, gemm_amd64.go): floats.Dot is FMA and lane-blocked like it; the change needs the tolerance its tests use and a measurement against the BLAS-backed np.dot
• Maximum/Minimum (pairwise, NaN-propagating; scalar Go in ndarray today)
• the ppc64le, loong64, riscv64 and s390x float64 kernels (see the go-simd/floats issue on porting them with NumPy's order)
• then delete ndarray's copies and its dead sum_<arch>.s; GEMM stays here (campaign rule: every user of dense matrix products already imports ndarray)