on runner startup, do all the ollama model warmup in one place

[lukemarsden]

• runners start multiple ollama processes
• ollama processes share the model cache on the filesystem in the container
• if requests result in multiple ollama processes warming up in parallel, they tread on eachothers' cache

I've seen issues with the model cache getting corrupted in this case. We should just warm up the ollama model cache once, based on the warmup models, before spinning up multiple ollama instances