IndexError in build_topology() when input mesh has fewer vertices than n_verts

Hi, thanks for releasing MeshFlow.

I encountered an `IndexError` when running `inference_vae.py` on a simple `.ply` mesh whose number of vertices is smaller than the model's configured `n_verts=4096`.

## Reproduction

I created a simple test sphere:

```python
import trimesh

mesh = trimesh.creation.icosphere(subdivisions=4, radius=1.0)
mesh.export("data/test/sphere.ply")
print(mesh.vertices.shape, mesh.faces.shape)
```

Then I ran:

```bash
CUDA_VISIBLE_DEVICES=0 python inference_vae.py \
--model_path ckpt/meshflow \
--input data/test/sphere.ply \
--output outputs/meshflow_vae/sphere_test \
--dtype bf16
```

## Error

```text
Loaded MeshFlowVAE from ckpt/meshflow
n_verts=4096 max_degree=50 point_feats_type=['embed@verts_normal', 'neighbor_points', 'embed@degree', 'embed@verts_mask']
sample_posterior=True dynamic_latent=True device=cuda dtype=bf16
Found 1 meshes in data/test/sphere.ply
Traceback (most recent call last):
File "inference_vae.py", line 168, in
main()
File "inference_vae.py", line 128, in main
topo = mesh.build_topology(
File "meshflow/utils/mesh.py", line 566, in build_topology
nearest_verts_idx = face_vertices[
IndexError: index 484 is out of bounds for dimension 1 with size 3
```

## Suspected cause

The error seems to happen in the padding branch when the input mesh has fewer vertices than `n_verts`.

In `meshflow/utils/mesh.py`, the following code appears to compute distances using face vertex indices instead of vertex coordinates:

```python
face_vertices = m.faces[negative_face_idx]
dist2 = (face_vertices - negative_verts[:, None, :]).pow(2).sum(dim=-1)
nearest_choice = dist2.argmin(dim=1)
nearest_verts_idx = face_vertices[
torch.arange(n_negative_verts, device=device),
nearest_choice,
]
```

Since `face_vertices` has shape `[N, 3]`, `nearest_choice` is expected to be in `{0, 1, 2}`. However, because the distance is computed against index values rather than 3D coordinates, `argmin` can produce invalid indices, which causes the out-of-bounds error.

## Temporary fix

The following patch fixed the issue on my side:

```python
face_vertices = m.faces[negative_face_idx]
face_vertex_pos = m.verts[face_vertices]

dist2 = (face_vertex_pos - negative_verts[:, None, :]).pow(2).sum(dim=-1)
nearest_choice = dist2.argmin(dim=1)

nearest_verts_idx = face_vertices[
torch.arange(n_negative_verts, device=device),
nearest_choice,
]
```

After this modification, `inference_vae.py` runs successfully and generates the reconstructed mesh.

Could you please confirm whether this is the intended fix, or whether the padding topology construction should be handled differently?

Thanks.