metal : promote mul_mv/mul_mm batch divisors to function constants

[guyfischman]

Overview

The batch-dimension fold i12 = im % args.ne12; i13 = im / args.ne12; and the GQA divisors i12 / args.r2, i13 / args.r3 execute on every dispatch of every mat-vec / mat-mat kernel. On Apple AGX, runtime int divmod is a ~700-cycle software path, while compile-time-known divisors fold to either a no-op (ne12=1), a shift (power-of-two divisor), or a magic-multiply (any other constant).

This PR promotes ne12, r2, r3 (and ne13 for mul_mm) to function constants so they are baked into the PSO at compile time:

FC_mul_mm_ne12/ne13/r2/r3 - bound by mul_mm pipeline from op shape
FC_mul_mv_ne12/r2/r3 - bound by mul_mv and mul_mv_ext pipelines from op shape; mul_mv_id binds to 1 since kernel_mul_mv_id wraps the impls with args0 = { ne12=1, r2=1, r3=1 }
The pipeline-cache key now includes ne12/r2/r3 so a Gemma-2 GQA PSO is not reused for a TinyLlama (ne12=1) op or vice versa.

This is the metal analog of #22650.

Additional information

Substituted args.ne12/r2/r3 with the FCs in 23 mul_mv impl sites, both mul_mv_ext impls, and both kernel_mul_mm variants (tensor and non-tensor branches).
Other kernels using args.ne12 (kernel_bin_fuse, kernel_soft_max, kernel_set_rows) are built by other pipelines that do not bind these FCs so they're unchanged.

The four FC values are bound as int16_t to match existing convention (nsg, nxpsg). Pipeline-cache call sites assert ne12/ne13/r2/r3 ≤ INT16_MAX before the cast to guard against silent wrap on edge cases like very large prompt batches. No-op for typical inference.

Correctness

./build/bin/test-backend-ops -b Metal
3/3 backends passed
Also verified bit-identicalness with llama-perplexity on all below models.

Performance

Apple M4 Pro, llama-bench -p 512 -n 128 -r 20, tg128 in t/s, master → patched:

TinyLlama Q4_0: 239.05 ± 1.73 → 246.57 ± 0.93 (+3.15%)
Llama 3.2 1B Q4_0: 227.58 ± 1.84 → 230.96 ± 2.65 (+1.49%, within noise)
Gemma 3 1B Q4_K_M (GQA): 164.67 ± 4.59 → 173.52 ± 0.75 (+5.37%)
Gemma 2 2B Q4_K_M (GQA): 103.41 ± 0.39 → 106.66 ± 0.24 (+3.14%)
Mistral 7B Q4_0 (GQA): 52.10 ± 0.08 → 52.73 ± 0.07 (+1.21%)
pp512 is unchanged across all five models (within ±1% noise). mul_mv is not on the prompt path; mul_mm shapes that hit the new FCs are too few to surface above bench noise.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Finding all instances of the antipattern, and looking for prior art (PR#22650)

[ggml-gh-bot]

Hi @guyfischman, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[guyfischman]

For reference I (using AI) made this PoC to see when literal vs FC wins and by how much over kernel arg/uniform buffers - https://github.com/SovereignSoft/agx-idiv-demo