2×2 SVD is ~7 digits less accurate through `DMatrix` than through `Matrix2`

```rust
use nalgebra::{Matrix2, DMatrix};

let vals = [2.27, -2.5e-6, 0.0, 9.008];
let a = Matrix2::new(vals[0], vals[1], vals[2], vals[3]);
let ad = DMatrix::from_row_slice(2, 2, &vals);

let e_static = (a.svd(true, true).recompose().unwrap() - a).norm();
let e_dynamic = (ad.clone().svd(true, true).recompose().unwrap() - ad).norm();
println!("static {e_static:.3e}");
println!("dynamic {e_dynamic:.3e}");
```

prints

```
static 2.051e-15
dynamic 8.966e-9
```

Reconstructing the same matrix from its own SVD (`U * Σ * Vᵀ`) is accurate to ~2e-15 through the static `Matrix2` path but only ~9e-9 through the dynamic `DMatrix` path — about seven digits worse for an ordinary, well-conditioned matrix.

Tested on nalgebra 0.35.0.

BTW, this bug was found using [hegel](https://crates.io/crates/hegeltest). Happy to contribute the tests if you're interested.

Start by reproducing the Matrix2 and DMatrix SVD reconstruction example against nalgebra 0.35.0, then trace the two SVD paths to find where their accuracy diverges. Done means the dynamic DMatrix reconstruction no longer loses roughly seven digits and regression tests cover the reported case.

Written by the indexing model from the issue text.