ROCm GGUF: split native JIT by operation family

[nekomario28]

Internal review/delivery carrier for the gfx1101-validated ROCm GGUF compile fix. Base is the existing ROCm + qwen35moe integration branch. CUDA keeps the original monolithic GGUF extension; ROCm lazily builds dequant, MMVQ, MMQ, MoE-vector and grouped-MoE modules. Physical RX 7800 XT evidence is recorded separately; this PR also includes the latest upstream FlashML-org#131 GGUF CPU/hybrid reachability commit with original author history preserved and a cheap Python routing contract test. Do not merge to fork main as a product release; use this as a review/delivery surface.

[nekomario28]

Physical evidence snapshot (RX 7800 XT / gfx1101, torch 2.11.0+rocm7.2, HIP 7.2.26015):

• the pre-split monolithic native GGUF TU HIPified with 0 unsupported CUDA calls but remained in active AMD-clang compilation beyond the bounded build window;
• production split dequant: cold JIT 16.006s, Q4_0 pure-Torch oracle PASS;
• production split MMVQ: cold JIT 16.426s, Q4_0 cosine 0.99995995 / NRMSE 0.008951;
• production split MMQ: cold JIT 115.677s, Q4_0 cosine 0.99993253 / NRMSE 0.011621;
• after replacing a HIPify-unsafe wrapper macro with the explicit switch in 93cc7ec, affected-only run 32940588112 / job 98090595262 passed: MoE-vector cold JIT 16.040s, Q4_0 cosine 0.99994171 / NRMSE 0.010843, IQ3_S cosine 0.99998331 / NRMSE 0.005793, grouped-MoE Q4_0 cosine 0.99997103 / NRMSE 0.007668;
• HIPify reported 0 unsupported CUDA calls in the successful follow-up.

Claim boundary: this proves the five native GGUF operation families compile/link/run numerically on physical gfx1101 through the split ROCm path. It does not yet claim full qwen35moe/Ornith model serving, all quant types numerically, or CUDA performance parity. CUDA remains on the existing monolithic path.

Persistent receipt: project-incubator research/2026-08-26-freetoken-rocm-gguf-operation-split-receipt.md.