feat: multi-model comparison benchmark for Crusoe Managed Inference

[Sakshi3027]

What this adds

A benchmark that runs multiple Crusoe models across reasoning, code generation,
and summarization tasks concurrently producing a latency and throughput leaderboard.

Models compared:

• meta-llama/Llama-3.3-70B-Instruct
• deepseek-ai/DeepSeek-V3-0324
• Qwen/Qwen3-235B-A22B

Tasks:

• Reasoning (multi-step math)
• Code generation (Sieve of Eratosthenes with type hints)
• Summarization (SQL vs NoSQL in 3 bullets)

All 9 combinations (3 models × 3 tasks) run concurrently via asyncio.gather.
Total benchmark time equals the slowest single call.

Why it's useful

Teams evaluating which Crusoe model to use for their workload need a
quick way to compare quality and speed across tasks. This gives them
a runnable starting point they can extend with their own prompts.

Testing

Tested locally with Groq (3 Llama variants) as a drop-in replacement.

To run on Crusoe:
export CRUSOE_API_KEY="your-api-key"
python compare.py

Related contributions

• feat: LangGraph multi-node research agent on Crusoe Managed Inference #55 — LangGraph multi-node agent
• feat: MLflow experiment tracking for Crusoe Managed Inference #56 — MLflow experiment tracking
• feat: RAG pipeline with Qdrant and sentence-transformers on Crusoe Managed Inference #57 — RAG pipeline with Qdrant
• feat: structured output and tool calling examples for Crusoe Managed Inference #58 — Structured output and tool calling
• feat: real-time streaming output examples for Crusoe Managed Inference #59 — Real-time streaming output
• feat: async batch inference benchmark for Crusoe Managed Inference #60 — Async batch inference benchmark