docs: point readers at akandr/bc250 for AI/LLM inference use case

The current docs frame the BC-250 mostly as a Linux gaming/desktop
board. akandr/bc250 is a 130-star community deep-dive on the AI
server use case (Ollama + Vulkan, 35B MoE at 37.5 tok/s, FLUX.2
image generation, 33-model benchmark suite, GTT + TTM memory
tuning, ROCm vs Vulkan analysis) which fills a gap our docs don't
cover and shouldn't duplicate.

Two minimal pointers added:

- index.md: one-line mention of the home-AI-server use case in the
  intro paragraph with a link to akandr
- 40cu-unlock.md: info admonition at the end of the thermal section
  noting that sustained LLM workloads have their own playbook and
  pointing at akandr for model selection, quantization, KV cache
  sizing, and long-context behaviour

No content duplication. Full credit to akandr, the goal is just to
help readers find the deep guide instead of reinventing it here.

docs/index.md

@@ -4,7 +4,7 @@ Documentation for the **AMD BC250** board — an ex-mining board featuring a cut
 
 ## What is the BC250?
 
-The AMD BC250 is a compact motherboard built around AMD's "Cyan Skillfish" APU, originally designed for cryptocurrency mining. The community has transformed this hardware into a capable Linux gaming and desktop system.
+The AMD BC250 is a compact motherboard built around AMD's "Cyan Skillfish" APU, originally designed for cryptocurrency mining. The community has transformed this hardware into a capable Linux gaming and desktop system, and more recently into a low-cost home AI server (Ollama + Vulkan inference up to 35B MoE models, see [akandr/bc250](https://github.com/akandr/bc250) for that use case).
 
 ### Key Specifications
 

docs/system/40cu-unlock.md

@@ -225,6 +225,9 @@ Sustained throughput drops about 10% over 10 minutes (3034 → 2835 tok/s) as th
 
 **Bottom line:** if you want to run a sustained overclock across all 40 CUs, you need effective cooling. Stock heatsinks and entry-level dual-fan setups will throttle under sustained load. With a 1500 MHz governor cap the heat stays manageable on the configuration above and the board runs comfortably. The unlock itself is solid: 25 minutes of looped Vulkan compute correctness testing produced zero `fp_errors`, zero `int_errors`, no amdgpu hangs, no resets, no oops. The thermal envelope is the constraint, not the silicon.
 
+!!!info "Running LLM / AI workloads on BC-250?"
+    [akandr/bc250](https://github.com/akandr/bc250) is a community deep-dive on running Ollama + Vulkan inference on this hardware (35B MoE at 37.5 tok/s, FLUX.2 image generation, 33 model benchmarks, GTT/TTM memory tuning, ROCm-vs-Vulkan analysis). If sustained AI inference is your use case, that guide covers everything this page intentionally skips (model selection, quantization impact, KV cache sizing, long-context behaviour).
+
 ## Selective CU Masking
 
 Not every unlocked CU may be silicon-healthy on every board. Boards with scattered (non-contiguous) harvest patterns may have defective CUs that pass enumeration but fail compute. duggasco ships a per-WGP health test that reboots into each WGP configuration in isolation and runs correctness checks: