Repin Inkling and kimi-k3 so the nightly merges onto b10289

[danielhanchen]

The nightly has failed since 08-05: ggml-org#25731 (Inkling) stopped merging once the base moved to b10280, and it stayed broken through b10285 and b10289.

Two pins move:

• ggml-org#25731: 1e6f9e4a59 -> 3fd7901cb4. The author merged master this morning, which resolves it upstream.
• unslothai#48: daef2b3e1b -> 522ef818d4. Our branch is stacked on Inkling, so the newer Inkling head conflicted with our older copy of it. I merged the new head into our branch; the only conflict was the WebGPU skip list, where our side is a superset (it adds MINIMAX_M3 and KIMI_K3 to their DEEPSEEK32/GLM_DSA), so the union is our side.

ggml-org#26185 deliberately stays at 04d6828b2b. Its newer head a614fab10b drops case LLM_ARCH_INKLING and KIMI_K3 from the WebGPU skip list, so it conflicts against the pins ahead of it. Newest is not always mergeable.

Verified by replaying the nightly's own merge sequence onto b10289 before opening this:

ok #24423 @ c3fb972412
ok #25731 @ 3fd7901cb4
ok #48   @ 522ef818d4
ok #26185 @ 04d6828b2b

[danielhanchen]

Superseded by #71, which pins #70 instead of #48 and adds the additive-merge fallback so this class of conflict resolves itself.