Migrate RL environments to Gymnasium - #53

Summary

• replace Gym 0.22 with Gymnasium 1.3 across RL environments, solvers, CI, installation, and documentation
• adopt Gymnasium reset and five-value step contracts while preserving Virne's non-RL simulation API
• keep termination and truncation distinct for correct GAE bootstrapping
• centralize step-result construction and rollout bootstrap handling
• add contract, fixed-trajectory, vector-environment, searcher, terminal, and truncation regression tests

Why

Gym 0.22 is obsolete and blocks dependency modernization. The migration removes the legacy package without changing deterministic environment or inference behavior.

Validation

• 256 passed
• Gymnasium check_env contract test
• fixed-seed action trace and GAE golden tests
• CPU ppo_dual_gat+ smoke: one training epoch, optimizer update, checkpoint save, and 4-VN inference
• strict Sphinx build: sphinx-build -W --keep-going
• pip check
• bash -n install.sh
• legacy gym package absent