Prebuilt: repin #24423 and #25731 onto carry branches

[danielhanchen]

The nightly has been failing in the merge loop since the aged base moved past b10549. Two third-party pins stopped merging:

• ggml-org#24423 (DiffusionGemma) conflicted in common/arg.cpp
• ggml-org#25731 (TML Inkling) conflicted in five files

The loop stops at the first conflict, so this did not just drop those two, it dropped the whole mix: #70, #91 and #95 all merge cleanly and were never applied.

Neither upstream PR can be repinned to a newer author commit. Both pins already equal their PR's current head, both are mergeable_state=dirty against upstream master too, and the authors have not touched them since 08-10 and 08-17. So both move to carry branches, following #99:

-  .../ggml-org/llama.cpp/pull/24423/commits/daca8075d871483545dd85d58ce11970b304b541
+  .../unslothai/llama.cpp/pull/107/commits/74acc40c37ae2eb36031981feda392b793944f72
-  .../ggml-org/llama.cpp/pull/25731/commits/c44c9a11a77cf47a24e69d79504bcbddf9f69017
+  .../unslothai/llama.cpp/pull/108/commits/60007bad6347a68bb662a2d490f20d383cad2678

The conflict resolutions and their verification are documented on #107 and #108. Both carries also needed one fix each for an upstream API change that did not conflict textually and so only showed up at build time: RESIZE_ALGO_BICUBIC_PILLOW was deleted upstream (#108), and the diffusion visual server had to move to the new common_json wrapper (#107). The second one matters here because the prebuilt workflows build the diffusion binaries best-effort and never fail the job, so it would have shipped bundles quietly missing them.

A side benefit: both pins are now fork-owned, so repin.py can maintain them automatically when the base tag moves, instead of emitting third-party and asking the author to merge master.

Verification

Replayed the merge loop from unsloth-prebuilt.yml against the aged base, in pr-set.json order:

base: b10630

CLEAN                    unslothai/llama.cpp#107  74acc40c37
CLEAN                    unslothai/llama.cpp#108  60007bad63
CLEAN (additive resolve) unslothai/llama.cpp#70   edfd4c1a3b
CLEAN                    unslothai/llama.cpp#91   c86ed26998
CLEAN                    unslothai/llama.cpp#95   3db8cb5b2e

Each pin was also probed in isolation onto the bare tag, so a clean result is not hiding an interaction between pins. All five are clean both ways.