Graphnetcore dataload compat

[JoKircher]

Summary

Brings the package up to GraphNetCore v0.4 and adds first-class data-import
converters (VTK and CSV) with a new user guide and a regression-tested suite.
The Documenter build is green and Aqua checks pass.

GraphNetCore v0.4 migration

• Depends on the registered GraphNetCore v0.4.1 (no git [sources]).
• Model/params/state now live in a Lux TrainState; parameter updates go through
  Setfield.@set! instead of field reassignment.
• Online-normaliser state is reconstructed from checkpoints; df_step schema widened
  and legacy 2-col checkpoints auto-migrated.
• Pins GPUCompiler to ~1.9 for Reactant/CUDA compatibility.
• Updated dataset.jl / graph.jl / solve.jl / strategies.jl, the example scripts,
  and test_normalizer.jl accordingly. GPU-verified on both fixtures.

Data import (new)

• vtk_to_hdf5 — converts ParaView / DualSPHysics VTK particle trajectories into the
  training HDF5 layout, one trajectory group per source:
  • XML .pvd / .pvtu / .vtu: the wrapper formats are parsed internally (ReadVTK can
    open neither), and only leaf .vtu pieces are handed to ReadVTK; empty (0-point)
    frames are detected and skipped before ReadVTK is called.
  • Legacy BINARY .vtk (POLYDATA, e.g. DualSPHysics PartAll_*.vtk) via a built-in
    reader — ReadVTK doesn't handle legacy .vtk.
  • Case-insensitive field mapping with alias fallback, id-based row alignment across time,
    recompute_acc (also automatic when no acceleration field is present), and either a
    user-provided or auto-generated (write_meta) meta.json.
  • New deps: ReadVTK, LightXML.
• csv_to_hdf5 refactored to share the on-disk writer (_write_trajectory!) with the
  VTK path so both stay byte-compatible.
• Fixed an always-throwing Dataset(datafile, metafile, args) constructor
  (format check used || where && was intended).

Docs

• New "Loading Data" guide covering every import path (directory/split loader, explicit
  Dataset, HDF5/JLD2, the meta.json schema, CSV, and VTK).
• Documented previously-undocumented exports (ModelConfig, save/load_model_config,
  update_meta!, visualize_eval) and fixed two README links that broke when the README is
  reused as the docs home — Documenter now builds with no warnings/errors.
• Project logo + favicon.

Tests

• test/test_converters.jl — CV1–CV11 (43 assertions), wired into runtests.jl. Fixtures are
  generated on the fly (WriteVTK for leaf .vtu; hand-written .pvd/.pvtu/empty/legacy .vtk)
  so the tests exercise our own parsing, empty-frame guard, and legacy reader. Also covers the
  csv_to_hdf5 round-trip.
• End-to-end verified against real data: WashTec .pvtu collection and a legacy .vtk
  trajectory (517 particles × 151 frames), both loading through Dataset/getobs.
• WriteVTK added as a test-only dependency.

Notes / follow-ups

• Variable particle count across frames is currently rejected (trim to a constant-count window).
• The skeleton meta.json only estimates default_connectivity_radius — set it explicitly
  for real training.
• Not yet covered: a .vtkhdf reader and guided meta.json creation.

🤖 Generated with Claude Code