Add fvdb.nn.Prune and generative shape completion / VAE examples

[swahtz]

Addresses #741.

Summary

• fvdb.nn.Prune — the MinkowskiPruning analog: prunes grid topology and its aligned features by a per-voxel boolean mask (pruned_grid + rmask, differentiable through kept rows). Registered in docs; 13 unit tests.
• examples/shape_completion.py / examples/shape_vae.py — one-to-one analogs of MinkowskiEngine's completion.py and vae.py, built on the generative transposed convolution from Unify sparse convolution and transpose semantics #726 (from_grid_batch_transposed(..., target_grid=None)) plus Prune.
• notebooks/04_shape_vae.ipynb — trains the VAE and explores its latent space: reconstructions, prior sampling, latent interpolation (ballet flat → boot), and an optional ipywidgets explorer. Shapes render as lit voxel surfaces.
• Dataset: the VAE trains on the 254-mesh GSO "Shoe" subset (CC-BY 4.0, Add Google Scanned Objects 'Shoe' subset (254 meshes, CC-BY 4.0) voxel-foundation/fvdb-example-data#1) via a new load_gso_shoes() loader and data-revision bump; completion uses the dragon + happy meshes.
• Removes examples/wip/structure_prediction_net.py (targeted a removed API surface; superseded by the new examples).

Design notes

• Per-level GT targets come from a conv_grid(2, 2) pyramid queried with coords_in_grid — not coarsened_grid, whose block-centroid lattice differs from what conv_transpose_grid(2, 2) inverts.
• ME's teacher forcing (keep |= target) and per-level BCE are reproduced verbatim; eval decodes without forcing.
• The VAE's dense-neck seed adds a learned per-voxel positional embedding — a uniform broadcast latent leaves neck voxels spatially indistinguishable after k2 transposes.
• Generative training is bound by per-iteration generated-topology grid construction (kernel maps are cheap); tracked upstream as Generated-topology grid construction dominates per-iteration ConvolutionPlan cost (serial per-grid build + merge) #755 rather than worked around here.

Results / verification

• Completion IoU ≈ 0.75 (500 iters, ~40 s); VAE reconstruction IoU ≈ 0.49 on a 254-shape dataset (1500 iters, ~3.5 min), prior samples decode to plausible shoes.
• pytest tests/unit/test_nn_modules.py: 81 passed (CPU+CUDA, float32/64).
• pytest tests/test_examples.py -k shape: 2 passed (~4 min, CI runpy path); notebook passes pytest --nbmake.
• black --line-length 120 clean.

🤖 Generated with Claude Code

[swahtz]

Added notebooks/04_shape_vae.ipynb: trains the shape VAE from examples/shape_vae.py and explores its latent space interactively — reconstructions (IoU 0.999), prior sampling, batched latent interpolation (car → dragon), and an ipywidgets latent explorer (blend two shapes + prior noise) with a static fallback when ipywidgets is unavailable. Executed outputs are stored so it renders on GitHub; passes pytest --nbmake in ~87 s on GPU. ipywidgets added to the fvdb_learn environment.

[swahtz]

Dataset upgrade: the shape VAE (example + notebook) now trains on the 254-model Google Scanned Objects Shoe subset (CC-BY 4.0, added to fvdb-example-data in voxel-foundation/fvdb-example-data#1) instead of the 6 bundled meshes — one category with real intra-class variation, mirroring the role of ModelNet40 chairs in ME's vae.py. Adds fvdb.utils.examples.load_gso_shoes() and bumps the pinned data revision; training runs random minibatches over the pre-voxelized GridBatch with a once-built conv_grid pyramid sub-indexed per batch. The notebook now interpolates a ballet flat into a tall boot.

CI note: pytest tests/test_examples.py -k shape now takes ~4:07 (was ~2:02) due to the larger dataset and minibatch training.

Perf note: generative training is bound by per-iteration generated-topology grid construction (conv_transpose_grid; the kernel maps themselves are cheap) — profiled and filed as #755 rather than worked around here, so the examples stay semantically 1:1 with MinkowskiEngine.

[Rishit-dagli]

are there any plans for a similar shape vae example but for 4d volume?