Improve model selector usability for endpoints with many models

[GO1984]

Summary

The dashboard model selector becomes hard to use when a provider exposes many models or when multiple endpoints are configured.

A useful improvement would be to group models by provider/endpoint and allow those groups to be collapsed or expanded.

Problem

When BenchLoop detects several providers or endpoints, the model dropdown can become long and difficult to scan.

This is especially noticeable with setups such as:

• Ollama with many local models
• OpenAI-compatible proxy endpoints
• multiple tunneled/local endpoints
• providers that expose many model IDs
• duplicate model names available from different endpoints

In those cases, users need to distinguish not only the model name, but also which endpoint/provider it belongs to.

Suggested UX

Group models by provider/endpoint, for example:

• Ollama (local)
• llama.cpp @ :8088
• vLLM @ :8000
• OpenRouter
• Custom endpoints

Each group could be collapsible. The selected model should still show both:

• model name
• endpoint/provider

Why this helps

• Reduces visual noise in the benchmark model selector.
• Makes duplicate model names across endpoints easier to distinguish.
• Helps users avoid accidentally benchmarking the wrong endpoint.
• Scales better for users with many local and remote models.

Notes

This should be a frontend UX change only. It should not affect benchmark scoring or provider behavior.