feat: add Ollama as runtime backend for Metal agent

[Defilan]

Summary

Adds Ollama as a third runtime option for the Metal agent (--runtime ollama) alongside the existing llama-server and omlx backends.

Ollama is the most widely adopted local LLM runtime (200K+ stars) and recently switched to MLX for Apple Silicon inference. Most Mac users already have it installed, making this the lowest-friction path to fast local inference with LLMKube.

Key advantages over the other backends:

• No model format changes needed (Ollama uses GGUF internally)
• No binary path management (Ollama manages itself)
• Model downloads handled by Ollama (/api/pull)
• Most users already have Ollama installed

How it works:

• OllamaExecutor manages models through Ollama's REST API
• Model pull: POST /api/pull with 5-min timeout for large downloads
• Pre-load: POST /api/generate with empty prompt (blocks until loaded)
• Readiness: GET /api/ps verifies model is in memory
• Unload: POST /api/generate with keep_alive: 0
• Includes model name mapping for all 16 catalog models (e.g., llama-3.2-3b -> llama3.2:3b)

Usage:

# Ensure Ollama is running
ollama serve

# Start Metal agent with Ollama runtime
llmkube-metal-agent --runtime ollama

# Deploy as usual
llmkube deploy llama-3.2-3b --gpu --accelerator metal

No CRD changes. Default runtime remains llama-server. Fully backward compatible.

Test plan

• make test passes
• Start agent with --runtime ollama, deploy model, verify inference
• Verify model pull works for a model not yet downloaded
• Delete InferenceService, verify model unloads (/api/ps empty)
• Default --runtime llama-server unchanged
• --runtime omlx still works

Runtime comparison

	llama-server	oMLX	Ollama
Model format	GGUF	MLX	GGUF
Model download	Manual/init container	Manual	Automatic (/api/pull)
Install base	llama.cpp users	Small	200K+ users
CRD changes	None	MLX format added	None
Process model	One per model	One daemon	One daemon

Builds on #257 (ProcessExecutor interface). Closes #248 (MLX backend goal achieved through both oMLX and Ollama paths).