Prebuilt: pin the IQ1_XS, IQ1_XXS and IQ1_XXXS quant types

[danielhanchen]

Adds the three sub-IQ1_S quantization types to the nightly mix: IQ1_XS at 1.4375 bpw, IQ1_XXS at 1.3125 and IQ1_XXXS at 1.1875.

Stacked on #90, so it targets that branch rather than master. #90 drops the merged 26841 pin, which currently fails the merge before anything after it is reached. Merge #90 first and this retargets to master on its own.

Pinning #91, not #61

The code is #61. The pin points at #91 because #61 branches off master past the squash sync 53fd974b4, and master shares history with upstream only up to 2026-06-10. resolve merges pins onto the upstream base tag, so #61 arrives 765 commits stale and conflicts across tools/ui/, vendor/cpp-httplib/ and vendor/sheredom/, files it never touches. #91 is the same commit on a b10356 base, the same arrangement #70 uses. #61 stays open for design review.

Resolve replay on b10356

including ggml-org/llama.cpp#24423 @ daca8075d871483545dd85d58ce11970b304b541
including ggml-org/llama.cpp#25731 @ 0a9841fa63edce1cd252ed6efea713715abb0cf2
including unslothai/llama.cpp#70 @ 06d2326acbf515b10f8d6abeada09123d361acde
including unslothai/llama.cpp#91 @ c86ed269986f2dced6325c5c58bda966a2e2ead1
warning: unslothai/llama.cpp#70 needed an additive merge; every conflict was a pure add/add
MERGED OK
  tag          b10356-mix-1c79028
  ggml tree    92d6a3e7335fd106486f466ab50dbdbd75d09971 (version 0.19.0)
  build number 10356

Only #70 needs additive_merge.py, and only for pure add/add hunks, same as today. The cmake/build-info.cmake stamping step still finds its three anchor lines in b10356.

Build and tests

Full mix built with CUDA sm_100 on a B200. The tree resolve produced hashes to cdc8f9490e4e33bc0dc87c94842455f6a8e5e189, identical to the tree these numbers come from.

check	result
build	completes, no errors
ctest	59 of 60; the 60th is test-backend-ops hitting the 1500s ctest timeout, run separately below
test-backend-ops -o MUL_MAT -b CUDA0	1219/1219, 2/2 backends
test-backend-ops -o MUL_MAT_ID -b CUDA0	874/874, 2/2 backends
quantize and generate, all three new types	works on Llama-3.2-1B with an imatrix

Sizes come out in the order the bpw table predicts: IQ1_S 393551840, IQ1_XS 380182496, IQ1_XXS 366813152, IQ1_XXXS 353443808.

Nothing else in the mix regresses

Built the same tree without the new commit and requantized from the same imatrix with both binaries. Byte-identical output:

IQ1_S    IDENTICAL  02dc2c2c1a63817dfb3998acc272923d8f68ed1ac10d06dc00e87b137c2c4482
IQ2_XXS  IDENTICAL  3a461ee80e793e93ab28b742da8513a9bd2342c5270f403a9398a7871efd657c
Q4_K_M   IDENTICAL  574d6de2b69c35bcb539dd210adee181b8eba1b7ae8864f525ef20b6ef83c2b3
Q8_0     IDENTICAL  05085ca8b9697115aeb2be01ee32873d476dec5b060abaaf2367b394fae82392

DiffusionGemma, Inkling, Kimi-K3 and Muse Glimmer all still land. Every arch registers in llama-arch.cpp, test-llama-archs passes, the four llama-diffusion-* binaries link, and the libraries export what they should (libllama.so: diffusion_gemma 17, inkling 24, kimi_k3 23, muse_glimmer 17; libmtmd.so: kimivl 7, muse_glimmer 14). Muse Glimmer comes from the base tag now that its pin is gone, at upstream's final state.