SGLang ROCm-specific configuration and AITER integration for AMD GPUs

SGLang on ROCm uses AITER as the default kernel backend in its Docker images. The AITER integration for DeepSeek-R1 replaces: (1) Complete MoE implementation (Top-K routing, MoE sorting, blockscale FP8 MoE computation), (2) FP8 blockscale GEMM for linear layers, (3) MLA decode attention and MHA prefill attention, (4) Custom all-reduce. Measured on MI300X with SGLang v0.4.4 + AITER: DeepSeek-R1 total throughput 14,635.98 token/s (median 2,255.85 token/s per decode step). The integration is based on a private repository (BruceXcluding/sglang tree aiter_customer_v0.4.4). SGLang's ROCm support is production-ready and listed in ROCm's supported frameworks. The key advantage over vLLM is SGLang's native RadixAttention for prefix caching, which complements AITER's kernel optimizations. RESOLVED (2026-10-04 cross-reference repair): the bare part names in affected_by_hardware were expanded to their record ids - mi300x -> amd-instinct-mi300x, mi325x -> amd-instinct-mi325x, mi350x -> amd-instinct-mi350x, mi355x -> amd-instinct-mi355x. Each name matched exactly one accelerator record in data/accelerators/, so no variant was chosen between siblings (MI300X and MI325X are both CDNA3, MI350X/MI355X/MI350P all CDNA4) - the id is forced by the part name. This field is still a mixed bucket: it also carries the architecture codename values gfx942/gfx950, which are not ids and were left alone.