Prebuilt: repin #25731 and #70 onto post-squash heads

[danielhanchen]

The nightly has been failing in Resolve tag since upstream squash-merged Kimi K3, so no prebuilt has been produced. Two pins needed moving.

What was failing

ggml-org#25731 (TML Inkling) was pinned to 0a9841fa, which no longer merges onto the base tag once #24423 is applied before it. Two of its four conflicts were real:

refused tools/mtmd/clip-model.h: both sides add the same line(s), so this is one change made twice: };
refused tools/mtmd/clip.cpp: merge base is not empty, so at least one side edited existing text

unslothai#70 (Kimi K3 vision) was pinned to 06d2326a, whose history carries the pre-squash snapshot of ggml-org#26185. Upstream merged that PR on 2026-08-15 as squash ad1de39e, and a squash is not an ancestor of the pinned commit, so the merge re-applied the whole text side on top of a base that already had it:

refused src/llama-arch.cpp: both sides add the same line(s), so this is one change made twice: case LLM_ARCH_KIMI_K3:

This is exactly the case the _doc in this file warns about.

What changed

Both PRs have been reconciled with current upstream, so this only moves the pins:

PR	Old	New
ggml-org#25731	0a9841fa	c44c9a11
unslothai#70	06d2326a	edfd4c1a

#25731 merged current master in. #70 was rebased onto master, which drops the duplicated text-side work and leaves only our delta, +209/-3 across 9 files: the MoonViT-3d vision tower, the full-size loading fixes and the SSM_A_NOSCAN placement fix. conversion/base.py fell out of that delta because upstream took repack_mxfp4_blocks as a ModelBase static method.

#91 and #95 are unchanged. #91 stays listed: it is merged upstream but its pin is still not an ancestor of the base tag, so it is contributing content.

Verification

Replayed the Resolve tag merge loop locally against base b10456 with the real scripts/unsloth/additive_merge.py:

base: b10456 (f275595dd)
CLEAN      ggml-org/llama.cpp#24423 @ daca8075d871
CLEAN      ggml-org/llama.cpp#25731 @ c44c9a11a77c
ADDITIVE   unslothai/llama.cpp#70   @ edfd4c1a3b7a  [resolved: tools/mtmd/CMakeLists.txt]
CLEAN      unslothai/llama.cpp#91   @ c86ed269986f
CLEAN      unslothai/llama.cpp#95   @ 3db8cb5b2e9b
RESULT: all pins stacked cleanly onto b10456

Both reconciled branches were built with CUDA on B200 and tested before pinning. #25731: test-llama-archs, test-chat, test-flash-attn-bias, test-flash-attn-generic-hash all pass, and test-backend-ops -o FLASH_ATTN_EXT_BANDED reports 13/13 on CUDA0. #70: test-llama-archs and test-chat pass.

One note for whoever touches this next: the new #70 pin sits 14 commits past b10456, since it is rebased on master rather than on the tag. The mix therefore carries a little upstream code that has not been through the 6 hour aging window.