KVarN exact-tail: fail-closed GGML_ABORT on unsupported structured body instead of materialization fallback

[raufaser]

ggml/src/ggml-cuda/fattn-tail.cuh:713 aborts the process ("unsupported structured body in exact-tail attention") when the exact-tail path packs a body whose structured form uses views and the KVarN re-dispatch of the packed body fails. Aborting instead of degrading to the materialized route is the issue; filed separately per review request on PR #149.

Reproducer (from the PR #149 review thread):

GGML_KVARN_TEST_NKV_LADDER_ONLY=1 GGML_KVARN_TEST_LADDER_TAIL=128 <test-kvarn>

History observed on gfx1100 (RX 7900 XTX):

• v0.4.7 base (pre-change): aborts at the first D512 ladder case (n_kv=256). Pre-existing, not a PR Make KVarN usable on ROCm RDNA3 (toward #122: D256 race, portable prefill, f32 WMMA prototype) #149 regression.
• Mid-branch PR trees: identical abort, same place.
• PR Make KVarN usable on ROCm RDNA3 (toward #122: D256 race, portable prefill, f32 WMMA prototype) #149 head f7f3350: no longer reproduces. The full tail ladder passes for TAIL=128 and TAIL=256 across D128/256/512 x n_kv 256..8192 (D512 RMSE ~1e-5). Likely re-routed by the MIXED domain enable in the round-2 commits, but that is not proven; other geometries may still hit the fail-closed path.

Suggestion: where the packed body is unsupported, degrade to materialization rather than GGML_ABORT. Keeping this open for the maintainer to decide whether the remaining risk warrants the fallback or the issue can close with the PR.

[Anbeeld]

Pushed a fix into v0.4.7 branch, please try it.