Upgrade dependency stack - #54

Summary

• move the supported baseline to Python 3.12/3.13 and add a reproducible uv.lock
• upgrade PyTorch, PyG, NumPy, SciPy, pandas, NetworkX, Gymnasium, and the remaining runtime/tooling dependencies
• modernize the CPU/CUDA installer, CI, packaging metadata, and Read the Docs environment
• preserve behavior across NetworkX, Hydra, and Python 3.12 API changes

Why

The previous dependency stack and installation path were increasingly difficult to install on current Python and accelerator environments. This update establishes a current, locked baseline while retaining the benchmark's existing runtime behavior.

Validation

• full test suite: 257 passed on Python 3.12 and Python 3.13
• clean CPU installation through install.sh, including pip check
• deterministic nrm_rank comparison: all 31 non-runtime result fields exactly equal across the old and new stacks
• ppo_dual_gat+ smoke training and inference: one epoch completed and 10/10 inference requests succeeded
• strict Sphinx build passed in both the full environment and a docs-only environment
• Ruff, warning-as-error compile checks, uv lock --check, shell syntax check, and source/wheel builds passed

Notes

• CUDA 12.6, 13.0, and 13.2 package indexes were verified, but GPU installation paths could not be executed on the local macOS host.
• PyG's PyTorch 2.13 wheel index currently publishes pyg_lib as the optional compiled extension; the complete test suite passes with the supported pure/PyG paths.