Prebuilt: pin the IQ1_XS, IQ1_XXS and IQ1_XXXS quant types onto master

[danielhanchen]

#92 was merged into its stacked base branch prset-drop-merged-26841 rather than into master, so the pin commit is sitting on that branch and never reached the scripts/unsloth/pr-set.json the nightly actually reads. Master today has only the #90 change:

$ git show origin/master:scripts/unsloth/pr-set.json | jq -r '.prs[]'
https://github.com/ggml-org/llama.cpp/pull/24423/commits/daca8075d871483545dd85d58ce11970b304b541
https://github.com/ggml-org/llama.cpp/pull/25731/commits/0a9841fa63edce1cd252ed6efea713715abb0cf2
https://github.com/unslothai/llama.cpp/pull/70/commits/06d2326acbf515b10f8d6abeada09123d361acde

This is the same one-line addition from #92, reapplied on top of master. Opened as a fresh branch rather than retargeting prset-drop-merged-26841, because that branch carries its own copy of the #90 commit and master took #90 as a squash, so pointing it at master would put a duplicate of #90 into the diff.

The pin still works now that #91 is merged

#91 merged into iq1-narrow-upstream-base, not into master, so the tree resolve builds (upstream base tag plus pins) still does not contain the change. The pin is what puts it there.

The gate handles this: a required pin that is no longer open keeps being merged in and logs a warning, which is exactly the case the _doc describes. The membership check also still passes, c86ed269 is a commit of #91. Once a base tag ever contains this work the entry should be deleted outright rather than left to no-op, per the note added in #90.

Verification

Unchanged from #92, since the pinned commit and the rest of the set are the same. Resolve replay on b10356 gives MERGED OK at tag b10356-mix-1c79028, with only #70 needing additive_merge.py and only for pure add/add hunks. The tree that comes out hashes to cdc8f9490e4e33bc0dc87c94842455f6a8e5e189, which is the tree that was built and tested:

• CUDA sm_100 build completes clean
• test-backend-ops -o MUL_MAT -b CUDA0 1219/1219, -o MUL_MAT_ID 874/874
• 59 of 60 ctest pass, the 60th being test-backend-ops itself hitting the 1500s ctest timeout
• all three types quantize and generate on Llama-3.2-1B with an imatrix, sizes in bpw order
• IQ1_S, IQ2_XXS, Q4_K_M and Q8_0 come out byte-identical against a build without the change
• DiffusionGemma, Inkling, Kimi-K3 and Muse Glimmer all still land in the merged tree