Community fixes 20260601

[GO1984]

Summary

This PR fixes several runtime issues found while using BenchLoop with local OpenAI-compatible servers and newer NVIDIA hardware.

Changes

• Detect OpenAI-compatible endpoints by common ports and hosts.
• Use /v1/chat/completions for OpenAI-compatible preflight checks.
• Skip Ollama version checks for OpenAI-compatible endpoints.
• Add endpoint-specific API key support via BENCHLOOP_OPENAI_KEYS.
• Forward endpoint-specific auth headers through model listing, chat, and streaming calls.
• Tolerate [N/A] values from nvidia-smi.
• Omit internal asyncio task objects from active run API responses.

Why

Some OpenAI-compatible servers, such as llama.cpp, expose:

• /v1/models
• /v1/chat/completions

but do not expose Ollama routes like:

• /api/tags
• /api/chat

Before this change, BenchLoop could treat those endpoints as Ollama and fail with errors like:

Health check failed (404): {"error":{"message":"File Not Found","type":"not_found_error","code":404}}