llama-swap raises queueing, AMD concerns

llama-swap looks more like a model lifecycle controller than an admission controller. If queueing matters, the safest pattern is to put backpressure above it and let the swapper stay focused on loading the right model.

• –The README centers on on-demand model switching, `groups`, and `ttl`, which help with residency and multi-model packing, but do not describe built-in request queuing.
• –An open feature request about preventing swaps while commands are running suggests in-flight work is still an edge case the project has to manage carefully.
• –AMD itself is not a blocker, but the repo ships a `vulkan` image and has an AMD Radeon Vulkan bug report around Docker UID/GID changes, so test your exact backend/container combo.
• –For a classroom API, a gateway queue or rate limiter in front of llama-swap is more predictable than hoping swaps will absorb contention.