feat(bench): add gbash and gbash-server benchmark runners

[chaliy]

Summary

• Add gbash subprocess runner (cold start, new Go process per run)
• Add gbash-server persistent child runner (JSON-RPC server on Unix socket, warm interpreter)
• Python wrapper script bridges gbash's JSON-RPC protocol to the PersistentChild stdin/stdout JSON lines protocol

Why

gbash is a Go-based sandboxed bash runtime for AI agents. Adding it to the benchmark suite enables direct performance comparison across all three runtimes (bashkit/Rust, gbash/Go, just-bash/TypeScript) in both subprocess and persistent execution models.

Smoke test results

Runner	startup_echo (ms)	Execution model
bashkit	0.086	in-process
gbash-server	1.86	persistent child
bash	9.14	subprocess
gbash	17.58	subprocess

• 96 benchmarks tested across all categories
• 100% output match on all supported commands
• Expected errors on awk (6) and jq (5) cases — gbash sandboxes unknown commands (exit 127)

Test plan

• cargo test --all-features — all 2564+ tests pass (pre-existing ssh_supabase failure unrelated)
• cargo fmt --check clean
• cargo clippy -p bashkit-bench clean (pre-existing warnings in bashkit lib only)
• ruff check and ruff format --check on Python wrapper
• Smoke tested both runners across startup, variables, arithmetic, control, strings, arrays, pipes, tools, complex, large categories