feat: add Ollama runtime support

[stamsam]

Summary

• add an Ollama runtime that streams through Ollama’s OpenAI-compatible chat completions endpoint
• add provider-aware model metadata and model picker labels
• discover locally installed Ollama models and surface them in the app
• document how to use Ollama models locally

Depends on #8 for the MLX port-collision fix. This PR is easiest to review after #8 lands; until then GitHub may show the shared port-selection commits in this diff.

Verification

• npm run typecheck
• npm run build
• curl http://127.0.0.1:11434/v1/models
• streamed a test response from http://127.0.0.1:11434/v1/chat/completions with gemma4:e2b

[rickmanelius]

Damn it. Posted my review to the wrong area. See feedback on original issue #7 (comment)

[stamsam]

Updated the draft with a provider-aware first-run setup flow:

• first boot now preselects an installed Ollama model when one is detected
• setup screen has explicit Ollama / MLX choices instead of a flat download-first model list
• installed Ollama models are labeled as installed and shown before suggested Ollama models
• MLX downloads remain available from the MLX tab

Re-verified with npm run typecheck and npm run build.

[stamsam]

Updated this draft to center oMLX instead of treating Ollama as the main external runtime:

• renamed the backend integration to external OpenAI-compatible runtimes
• added first-class oMLX discovery from http://127.0.0.1:8000/v1/models
• added OMLX_HOST and OMLX_API_KEY support
• first-run setup now has oMLX / MLX / Ollama choices
• oMLX models discovered from the server can be selected directly without downloading managed MLX weights

I also built a local packaged app with npm run pack.

[stamsam]

Added setup-time oMLX key entry and model refresh:

• users can paste an oMLX API key in first-run setup
• the key is applied in the main process immediately, without restarting
• Refresh re-queries /v1/models and updates the model list
• empty oMLX state now tells users to enter a key if needed and refresh

Rebuilt the local packaged app with npm run pack.