Board-verify the FunctionGemma KV-cache decode loop on the SL2610

[michalharakal]

The KV-cache 2-graph decode (GemmaKvDecoder, GEMMA_KV=1) is CPU-verified (transformers PR #245) but the
board loop is an UNVERIFIED draft. On the first SL2610 run, confirm each silent-corruption trap:

• Per-block K-vs-V output order — GemmaKvDecoder.kFirstInOutput=false ((V,K) per return-SSA); flip to true if the first run yields garbage tokens.
• --output=@file format — IreeRuntime.invokeFiles assumes RAW bytes; strip the .npy header on read if the board writes NumPy.
• gemma_with_past input arg order — trace-derived; confirm against the compiled vmfb.
• Dynamic-concat — the x?x256 sentinel-relax leaves shape inference to iree-compile; confirm the vmfb accepts it (else fixed-pad+mask fallback).
• --task_topology_group_count accepted by the board iree-run-module (gated to revert).
• vmfbs built with the Torq-fork iree-compile (g165e12a).

Done: the loop reproduces the oracle [262146,236769,3255,718,498,1373,262152,106], PERF-LOGBOOK.md shows the O(seq²)→O(seq) collapse vs ~6000 ms/token, then make GEMMA_KV=1 the default in Pipeline.kt.

Runbook: sl2610-function-calling/docs/GEMMA-KV-BOARD-LOOP.md. Tracker: docs/GEMMA-KV-INT8.md.