PR 10: Add refactor CI and recovery benchmarks

[insujang]

Summary

• adds Python lint, formatting, Pyright, Rust, CPU, and CUDA-over-Gloo CI gates
• adds machine-readable balanced recovery scheduling results
• adds steady-state overhead and transactional replay benchmark results
• validates numerical equivalence and source-load reduction

Validation

• benchmark tests: 2 passed
• Ruff and formatting passed
• Pyright: 0 errors, 0 warnings
• cargo test --locked -q: 6 passed
• workflow YAML and diff checks passed

Targets only refactor.