[cute_math] Transpose issue with m3

[hgarrereyn]

Hi, it seems like there is some inconsistency with how rows/cols are interpreted in cute::m3.

Specifically, a property about outer_product consistency fails to hold:

Given three $n \times 1$ vectors: $u$, $v$, and $x$, the following equality should hold:

$$ (u \times v) w = (v \cdot w) u $$

However, in cute_math, the outer product of $u \times v$ actually computes $v \times u$ (transposed).

It's not entirely clear which way this should be fixed, because outer_product constructs the m3 with rows(...) (which is correct if these are matrix rows), however cute::mul interprets those as columns, hence the transposition.

(found via automated fuzzing)

The issue is demonstrated in the following testcase:
testcase.cpp

#include <cmath>
#include <cassert>
#include <cstdio>
#include "cute_math.h"

int main()
{
    // Small, deterministic values to show the error clearly.
    cute::v3 u(1.0f, 2.0f, 3.0f);
    cute::v3 v(4.0f, 5.0f, 6.0f);
    cute::v3 x(7.0f, 8.0f, 9.0f);

    cute::m3 M = cute::outer_product(u, v); // Expected: M = u v^T
    cute::v3 lhs = cute::mul(M, x);

    printf("lhs: %f, %f, %f\n", cute::getx(lhs), cute::gety(lhs), cute::getz(lhs));

    // Property for true outer product: M*x == u * dot(v, x)
    float s = cute::dot(v, x);
    cute::v3 rhs = u * s;

    printf("rhs: %f, %f, %f\n", cute::getx(rhs), cute::gety(rhs), cute::getz(rhs));

    float dx = cute::getx(lhs) - cute::getx(rhs);
    float dy = cute::gety(lhs) - cute::gety(rhs);
    float dz = cute::getz(lhs) - cute::getz(rhs);

    // This assertion fails with the current implementation because
    // outer_product appears to return v u^T instead of u v^T.
    float tol = 1e-5f;
    assert(std::fabs(dx) <= tol && std::fabs(dy) <= tol && std::fabs(dz) <= tol);
    return 0;
}

output

lhs: 200.000000, 250.000000, 300.000000
rhs: 122.000000, 244.000000, 366.000000