GraphZero v0.2 brings zero-copy GNN training via mmap

The "OOM wall" in GNN training is a real and underserved problem — PyG and DGL assume you have the RAM, and if you don't, you crash. GraphZero's mmap approach is the right systems-level answer.

• –Zero-copy via mmap means OS page faults do the heavy lifting — only touched 4KB blocks land in RAM, dropping peak usage from 24GB+ to ~5GB on ogbn-papers100M
• –Custom CSR binary format (`.gl`) achieves ~60% compression over raw CSV topology; columnar `.gd` maps directly to PyTorch tensors with zero allocation overhead
• –OpenMP neighbor sampling releases the Python GIL, enabling true parallelism across disk I/O, CPU sampling, and GPU compute
• –Currently scoped to data loading and sampling, not training itself — a deliberate narrow focus that keeps the library composable with any training loop
• –Very early (7 GitHub stars, v0.2), but the roadmap toward dynamic graph updates and ACID multi-process safety signals serious long-term ambition