Perf: Inside column-major GEMV, reduction can be optimized using SIMD loads and FMA

[SwayamInSync]

Current implementation:

for (size_t i = 0; i < m; ++i) {
    y[i * incy] = Sleef_fmaq1_u05(col[i], x_j, y[i * incy]);
}

can be parallelised as following given the vector stride is 1 otherwise we can fallback to the original

if (incy == 1 && m >= VECTOR_SIZE) {
    // Vectorized path for unit stride
    for (size_t i = 0; i < m - (m % VECTOR_SIZE); i += VECTOR_SIZE) {
        QuadVector y_vec = QuadVector::load(&y[i]);
        QuadVector col_vec = QuadVector::load(&col[i]);
        QuadVector x_j_vec(x_j);
        
        y_vec = col_vec.fma(x_j_vec, y_vec);
        y_vec.store(&y[i]);
    }
    
    // Handle remainder
    for (size_t i = m - (m % VECTOR_SIZE); i < m; ++i) {
        y[i] = Sleef_fmaq1_u05(col[i], x_j, y[i]);
    }
} else {
    // Current scalar path for strided access
    for (size_t i = 0; i < m; ++i) {
        y[i * incy] = Sleef_fmaq1_u05(col[i], x_j, y[i * incy]);
    }
}