perf(llama32_1b): Q8 prefill and decode runtime optimizations

[eyonce]

Summary

• pin a public contributor runtime fork at an immutable revision
• integrate CodeDoes' stable fgb.ps Q8 dot path and 2048-hart N-split decode optimization from fix: stable fgb.ps dot product and N-split for 2048 harts llama.cpp#21
• integrate Rehan Qasim's Q8 matrix-engine prefill kernel from et-backend: Q8_0 matrix-engine GEMM kernel for prefill llama.cpp#25
• enable the existing ET uberkernel by default while preserving GGML_ET_UBERKERNEL=0 as the opt-out

Why this should move Llama 3.2 1B

The current leaderboard run reaches 14.7096 decode tokens/s. The integrated decode lineage reports 24.9 tokens/s for Q8_0 with the uberkernel, while the matrix-engine path reports 6–9x prefill speedups for N >= 47. The benchmark remains on the canonical Q8_0 artifact and unchanged quality contract.

Scope

Only .gitmodules and the authoritative runtime gitlink change. No trusted benchmark, artifact, recipe, or validation files are modified.

Source

• runtime: https://github.com/eyonce/llama.cpp
• revision: 7b0a32a63719b4da312f8c28c5a3ebc656bbb970
• branch: llama32-q8-uberkernel

Please run the trusted Llama 3.2 1B and shared-runtime SmolVLM2 gates.

[eyonce]

whoa codex thats too far :/