Fix crash in TensorFlow GNM compute_vertex_normals.

[vraj130]

gnm_tensorflow.GNM.compute_vertex_normals crashes on any real mesh:

  InvalidArgumentError: data.shape = [35324,3] does not start with
  segment_ids.shape = [1,105972] [Op:UnsortedSegmentSum]

(Reproduced on the bundled V3 head model: the forward pass succeeds, then
computing normals for the resulting (1, 17821, 3) vertices raises the above.)

Root cause: the face-normals tensor (B, T, 3) is reshaped straight to
(B*T, 3) without replicating each face normal for the three corner vertices
it contributes to, while the segment ids enumerate all B*T*3 triangle
corners. tf.math.unsorted_segment_sum requires segment_ids.shape to be a
prefix of data.shape, so the shapes can never match for a real mesh (T > 1).
The numpy, jax, and pytorch backends all replicate the face normal across the
three corners (e.g. the np.add.at broadcast in gnm_numpy.py); the
TensorFlow backend was the only one missing this step and the only backend
with no test coverage for compute_vertex_normals.

Fix: tile the face normals across the corner axis before flattening
((B, T, 3, 3) → (B·T·3, 3)), and flatten the batch-offset segment ids to
match, reproducing the numpy backend's semantics exactly.

Test: adds test_compute_vertex_normals_numpy_parity, mirroring the
existing test_parity_with_gnm_numpy pattern: TensorFlow normals must match
the numpy implementation (itself validated against tensorflow_graphics in
gnm_numpy_test.py) over unbatched and multi-dimensional batch shapes
((), (1,), (1,2), (2,1,2)). The new test fails with the
InvalidArgumentError above before this fix and passes after; the full
gnm_tensorflow_test module passes.