kyuz0 amd-strix-halo-toolboxes building.md retained-PM4 measured llama-server flags

docs/building.md retained-PM4 Strix Halo build (rocm-10.0-strix-llama). Measured 1207 t/s prompt processing (pp2048, depth 0) and 43.9 t/s decode on Qwen3.8-Flash-Next with its shared MTP head. llama-server flags: -dev ROCm0 -ngl all -fa on -fit off --load-mode none --lazy-mode on-direct -ctk f16 -ctv f16 -c 262144 -b 16384 -ub 16384 --parallel 1 --spec-type draft-mtp,ngram-mod --spec-draft-n-max 3. Engine halo-box/strix-llama.cpp, custom ROCm runtime from pwilkin/rocm-systems:ilintar-experiments. No llama.cpp commit hash and no host RAM size are stated.