[Proposal] Optimization for AoS with SIMD

Not sure where this should be best done - at the compiler level, at the binaryen level, or at the VM runtime level.

As the GC proposal adds structures and arrays, we technically have the ability to turn AoS into SoA or AoSoA (with fixed SIMD proposal).
Like (pseudocode):

class Vec3 {
    f32 x;
    f32 y;
    f32 z;
};
const vels = Vec3[100];
f32 getVelX(i32 i) { 
  return vels[i].x;
}

into:

class Vec3 {
    f32 x[4];
    f32 y[4];
    f32 z[4];
};
const vels = Vec3[(100 + 3) / 4];
f32 getVelX(i32 i) { 
  return vels[i / 4].x[i & 3];
 }

And apply SIMD auto-vectorisation afterwards.

No file, test, or entry point is identified. Start by determining whether the transformation belongs in the compiler, Binaryen, or VM runtime, then define representative AoS inputs and a SIMD benchmark. Done means a chosen layer, a semantics-preserving AoS-to-SoA or AoSoA transformation, and demonstrated auto-vectorization.