feat: add SGLang local model serving to comparison framework (#190)

Add support for serving models via SGLang on remote GPU hosts, enabling
comparison of API models (GPT, Claude) against locally-served models
(e.g. Qwen3.5-9B) that vLLM cannot serve.

Key changes:
- New scripts/sglang_server.py: SGLangServerManager handles full lifecycle
  (SSH install, server start, readiness polling, SSH tunnel, cleanup)
- Extended ModelConfig with provider="sglang", serve config, max_new_tokens
- New --gpu-host and --ssh-key CLI flags (optional; sglang models skipped
  without --gpu-host)
- SGLang server auto-starts per model, tunneled as OpenAI-compatible API
- Environment variables (OPENAI_BASE_URL, OPENAI_API_KEY) saved/restored
  between models so API models remain unaffected
- New example_comparisons/unified_agents.yaml demonstrating mixed config

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

Author: abrichr

example_comparisons/unified_agents.yaml

@@ -0,0 +1,41 @@
+# Example comparison config: API + local (SGLang) models as unified agents
+#
+# Usage:
+#   # API models only (no --gpu-host, sglang models are skipped):
+#   python scripts/compare_models.py --config example_comparisons/unified_agents.yaml
+#
+#   # API + local models (SGLang on remote GPU):
+#   python scripts/compare_models.py \
+#       --config example_comparisons/unified_agents.yaml \
+#       --gpu-host user@gpu-server
+#
+#   # With specific SSH key:
+#   python scripts/compare_models.py \
+#       --config example_comparisons/unified_agents.yaml \
+#       --gpu-host user@gpu-server --ssh-key ~/.ssh/gpu_key
+
+name: "Unified Agent Comparison (API + Local)"
+description: "Compare API models vs locally-served models as unified desktop agents"
+
+tasks:
+  - example_tasks/notepad-hello.yaml
+  - example_tasks/clear-browsing-data-chrome.yaml
+
+models:
+  - name: gpt-5.4-mini
+    provider: openai
+    type: unified
+
+  - name: Qwen/Qwen3.5-9B
+    provider: sglang
+    type: unified
+    max_new_tokens: 2048
+    serve:
+      engine: sglang
+      port: 8080
+
+server_url: http://localhost:5001
+max_steps: 10
+runs_per_config: 1
+save_screenshots: true
+output_dir: comparison_results/unified/