Add Ollama provider with model lifecycle management

[djthorpe]

Summary

Implement an Ollama provider (pkg/provider/ollama) for local model inference, including full model lifecycle management (download, load, unload, delete).

Requirements

Core Provider

• Implement llm.Client interface for the Ollama API (http://localhost:11434 by default)
• Endpoint configurable via OLLAMA_ENDPOINT environment variable or CLI flag
• No API key required (local service)
• Support streaming and non-streaming chat completions
• Support tool/function calling (Ollama supports this for compatible models)
• Support embeddings

Model Management

• Pull/Download: Pull models from the Ollama registry (equivalent to ollama pull)
• List: List locally available models with size, quantization, and modification date
• Load: Load a model into memory (warm up for faster inference)
• Unload: Unload a model from memory to free GPU/RAM
• Delete: Remove a model from local storage
• Show: Get model details (parameters, template, license, system prompt)
• Expose model management operations via the API (new endpoints or extend existing model API)

API Endpoints

• POST /api/model/pull — pull a model by name/tag
• POST /api/model/{name}/load — load model into memory
• POST /api/model/{name}/unload — unload model from memory
• DELETE /api/model/{name} — delete a model
• Pull progress should be streamable (SSE or chunked response) for UI feedback

Models

• Any model available in the Ollama registry (llama, mistral, gemma, phi, qwen, etc.)
• Model names use the Ollama format: model:tag (e.g. llama3.2:latest, mistral:7b-instruct-q4_0)

Notes

• Ollama API docs: https://github.com/ollama/ollama/blob/main/docs/api.md
• The existing ollama.tf deployment in tf-modules shows Ollama is already running in the infrastructure
• Consider how model management interacts with the CLI (llm model pull, llm model delete, etc.)

Motivation

Ollama enables local/private model inference with no API costs. Model management is a key differentiator — users need to download, load, and manage models on their GPU servers. This complements the cloud providers (Gemini, Anthropic, Mistral) with self-hosted options.

OpenAI is a major LLM provider and its multi-modal output capabilities (images, audio) will drive the content model to support rich responses across all providers, improving the overall architecture.