ai: add shimmy as alternative LLM backend

[awdemos]

Shimmy is a lightweight Rust-based OpenAI-compatible API server for GGUF models. It supports CPU and GPU inference (CUDA, Vulkan, OpenCL, MLX).

• Add shimmy component alongside existing ollama
• OpenWebUI and n8n can use either backend (ollama preferred, shimmy fallback)
• CPU by default it auto-detects GPU when enabled
• Health checks on /health endpoint

Users can choose between ollama or shimmy.

[AdamNowotny]

Thanks @awdemos ! So this can replace Ollama for both OpenWebUI and N8n?
Might be worth looking into putting ollamaUrl in the config soon so users can decide what they prefer even if both are installed...

Let's get this one merged first, can you just use image instead of imageTag? It's a recent change in preparation for being able to have a local image registry as well.

Also there is no need to add these AMD specific vars, it's handled in https://github.com/QC-Labs/orange-lab/blob/main/src/containers.ts#L175-L183

[AdamNowotny]

not needed anymore, just needs <app>:image (shimmy:image=ghcr.io/michael-a-kuykendall/shimmy:latest) added to Pulumi.yaml

[AdamNowotny]

standard AMD driver vars, handled in containers.ts whenever <app>:gpu=amd, can be removed

[AdamNowotny]

@awdemos did you get this working with ghcr.io/michael-a-kuykendall/shimmy:latest image? I had to add shimmy to command args to prevent the pod crashing on start but even then GPU support and llama backend is not included in that image:

$ shimmy gpu-info
🖥️  GPU Backend Information

❌ llama.cpp backend not available (compile with --features llama)
🍎 MLX Backend: Disabled (compile with --features mlx)

💡 To enable GPU acceleration:
   cargo install shimmy --features llama-cuda    # NVIDIA CUDA
   cargo install shimmy --features llama-vulkan  # Cross-platform Vulkan
   cargo install shimmy --features llama-opencl  # AMD/Intel OpenCL
   cargo install shimmy --features gpu           # All GPU backends

That causes all models fail to load, even with CPU.

[AdamNowotny]

New image has been published so this might be worth revisiting (Michael-A-Kuykendall/shimmy#146 (comment)). The files need to move to stacks/ai