feat: add GPU instance lifecycle CLI for model serving (#189)

Add `openadapt-gpu` CLI that automates the full GPU experiment workflow:
launch EC2 instance, install inference engine (SGLang/vLLM), serve model,
set up SSH tunnel, and terminate. Replaces 9 manual steps with one command.

Commands: launch, status, terminate, serve, run-comparison, ssh, logs.

State persisted to ~/.openadapt/gpu_state.json so terminate always works
even when SSH is broken (uses EC2 API directly).

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: abrichr

pyproject.toml

@@ -116,6 +116,8 @@ openadapt-train-grpo = "scripts.train_trl_grpo:main"
 openadapt-eval = "scripts.run_full_eval:main"
 openadapt-collect = "scripts.collect_distillation_data:main"
 openadapt-analyze = "openadapt_evals.analysis.cli:main"
+# GPU instance lifecycle (launch, serve, terminate)
+openadapt-gpu = "scripts.gpu_cli:main"
 
 [project.urls]
 Homepage = "https://github.com/OpenAdaptAI/openadapt-evals"