AMD MI100 multi-GPU segfaults on bare-metal ROCm - needs PCIe-passthrough VM + HSA_AMD_SVM

Adding a second (or more) MI100 card causes segfaults and instability on bare-metal Linux with standard ROCm - the canonical llama.cpp report (#2824) could only get 2x MI100 working by passing the cards through to a VM. Single-card works; multi-card is the landmine.

gfx908/CDNA1 is a maintenance-tier citizen on the ROCm compat matrix - multi-GPU shared virtual memory is fragile on bare metal. The working recipe requires the HSA_AMD_SVM kernel option and PCIe passthrough into an Ubuntu VM, which sidesteps the bare-metal driver path that segfaults.

Run each MI100 (or the multi-card set) inside an Ubuntu VM via PCIe passthrough, and set the HSA_AMD_SVM kernel option. Read llama.cpp discussion #2824 end-to-end BEFORE buying the second card - the measured 8.32 tok/s on Llama-2-70B Q6_K was achieved under exactly this setup, not bare metal.

MI100 is the best $/(GB x TB/s) in the used slice (~$720-1,299 for 32GB/1228GB/s) but the ROCm tax is the price. gfx908 is still on the ROCm 10 compat matrix but is maintenance-tier. gfx90a (MI210/MI250) does NOT have this problem - it is fully supported upstream with no fork needed.