feat: repack quantized CLIP weights into the CPU i8mm buffer

[olyasir]

Summary

Speeds up the CPU vision encoder (mtmd/clip) by repacking quantized weights
into ggml's CPU i8mm buffer. On a Pixel 9 Pro (Qwen3.5-0.8B mmproj, q8_0,
4 threads) the vision encode drops from ~1807 ms → ~1114 ms (1.6×), with
bit-identical output.

Production change only — one file (tools/mtmd/clip.cpp, +72/-2). The standalone
qwen3vl-encoder benchmark tool used to find/validate this lives separately on
branch feat/qwen3vl-standalone-encoder (PR #178) for testing.

What it does

clip.cpp allocated all weights in the default CPU buffer, so the q8_0/q4_0
matmul matrices used the generic GEMM instead of the interleaved i8mm/AVX2
kernels. This allocates the repackable weights in ggml's CPU "repack" extra
buffer type, so they are repacked into the interleaved layout at load.

• Per-tensor selection via a ggml_backend_dev_supports_op probe (mirrors
  llama's buft_supported): only tensors the backend can actually repack+matmul
  go to the repack buffer. Conv weights / biases / norms stay in the default
  buffer (otherwise their non-matmul ops become unsupported and the scheduler
  aborts). On arches/variants with no repack kernel the probe rejects everything
  and this is a no-op.
• Weight loading is made per-tensor host-aware, since weights now span the host
  default buffer and the non-host repack buffer.
• Bit-identical output (verified via cosine-similarity against a non-repack
  golden). Opt out with MTMD_CLIP_NO_REPACK=1.

Production android already ships an i8mm variant (GGML_CPU_ALL_VARIANTS →
armv8.6_1), so this repack is the incremental win on top with no build-config
change needed.

Testing

• Built x86-64 (Vulkan) and arm64-android (CPU/Vulkan/OpenCL).
• Repack path verified bit-identical to non-repack via cosine-similarity.
• On-device on Pixel 9 Pro (Mali, CPU path) via Firebase Test Lab.
• Wired end-to-end through the addon consumers via monorepo overlay
  (test[notask]: overlay qvac-fabric onto qwen3vl CPU-repack branch (fabric #178) qvac#3012).

[olyasir]

CPU vision-encode benchmark — repack ON vs OFF (same binary)

Measured this PR's repack in the qvac VLM benchmark (several-sources, cognitive preset, Qwen3.5-0.8B · Q8 mmproj), toggling the repack at runtime via MTMD_CLIP_NO_REPACK on a single build — so it's a clean same-binary A/B with no build/thermal confound.

mmproj vision-encode time (lower = faster):

Platform (CPU)	repack OFF	repack ON	speedup
Pixel 9 (Android ARM)	6117 ms	4380 ms	1.40× (−1737 ms)
S25 (Android ARM)	2532 ms	1812 ms	1.40× (−720 ms)
mac-mini M4 (ARM)	2400 ms	1431 ms	1.68×

• Lossless — VQA quality identical (46.7% on every leg, repack on == off).
• TTFT drops correspondingly (Pixel 10328 → 8089 ms) since it includes the encode.
• x86 legs are a no-op by design (q8_0 repack is ARM NEON + i8mm only).

Benchmark results: https://github.com/tetherto/qvac/actions/runs/28658081618/job/84996939580 (run 28658081618, matrix-combine → Summary)

_{Mobile encode times are captured via the visionEncodeMs runtime stat added in qvac#3028; the A/B used a monorepo overlay pinning this branch (a3bc3d8d).}