wasm: use relaxed-simd madd for v_muladd when available

[MMesch]

Summary

When emcc is invoked with -mrelaxed-simd, __wasm_relaxed_simd__ is defined and the wasm_f{32x4,64x2}_relaxed_madd intrinsics become available. These lower to hardware FMA on engines that support wasm relaxed-simd (Chrome >=114, Firefox >=120, Safari >=16.4, Node >=22) and to a separate mul+add otherwise, so callers get FMA where possible with no ABI change.

Off by default: builds that do not pass -mrelaxed-simd fall through to the existing separate mul+add definitions of v_muladd_f32 / v_muladd_f64, so no behavior change for the default wasm build.

Notes

• Change is scoped to kernel/simd/intrin_wasm.h (+8 lines).
• Level-3 wasm128 kernels (gemmkernel_4x4_wasm128.c, trmmkernel_4x4_wasm128.c, zgemmkernel_2x2_wasm128.c) already call wasm_f{32x4,64x2}_relaxed_madd directly under __wasm_relaxed_simd__. This patch extends the same behavior to everything that goes through v_muladd, so Level-1 kernels benefit too.
• Downstream verification (built with -msimd128 -mrelaxed-simd, disassembled with wasm2wat): 25 additional relaxed_madd emissions attributable to this patch, in sdsdot_k (10) and sdot_k / ddot_k / srot_k (5 each) via kernel/generic/dot.c.
• Preliminary benchmark: on a large-n numpy.correlate workload, wasm overhead vs. native drops from ~7x to ~3.5x on relaxed-simd-capable engines (roughly 2x speedup of the wasm build itself). More rigorous benchmarks needed to confirm.