metal : add PTQ1_0 mat-vec for two to four columns

[jasontitus]

Overview

Add an opt-in Metal PTQ1_0/F32 mat-vec for 2-4 columns, sharing decoded weights across columns and activations across four output rows. Enable with GGML_METAL_PTQ1_MULTICOL=1; default remains off. Existing single-vector and larger-batch paths are unchanged. Five files; no benchmark framework, MTP tooling or device-specific dispatch.

Single-user Bonsai 2 PTQ1_0 native generation tok/s:

Device	Original plain	Original MTP	Optimized plain	Optimized MTP	MTP gain over optimized plain
M1 Ultra	26.93	10.91	27.26	37.18	+36.4%
M5 Max	30.19	13.01	29.98	39.48	+31.7%

For an M5 single user, candidate MTP is about 32% faster generation than candidate plain. M5 full-request throughput, including prompt/request overhead, is 27.96 plain and 36.25 MTP. The 2.872x paired gain compares kernel on/off with MTP already enabled.

Additional information

Related: #218 (CUDA small batches), #225 (Metal single-vector work).

Measured base: 0324c66521960d67aa7da8687fb1453a79a6565c, research R4/S1. PR base: 0781925904391351963d499cb32cd735849b06a5. Same kernel arithmetic; unused geometries/tuning switches removed. Performance has not been remeasured on the new base.

Three ABBA quartets per cell, six observations per variant. Short prompts, up to 128 generated tokens; 4K context/slot, full Metal offload, Flash Attention, F16 KV, 16 CPU threads, batch/ubatch 512. Plain/MTP are separate randomized cells. MTP uses a separately grafted pinned Qwen3.8 head and one draft token; this patch does not provide MTP.

• Current-base M1: Release build passed; 172/172 CPU-reference backend cases and 42/42 numerical checks passed with selector off and on. Two-chunk WikiText-2 perplexity: 8.7487 off, 8.7486 on (context 512, batch/ubatch 4).
• Measured PTQ1_0 server output comparisons matched; all 168 M5 plain/MTP comparisons matched. Logits are not bitwise equal (M5 max absolute difference 0.0012482, max NMSE 3.003e-10).
• M1 completed standard llama-bench. M5 short-prefill quartets failed stability checks; no accepted M5 llama-bench result is claimed. Thermal samples were predominantly Heavy (M1 88.7%, M5 83.3%).
• Five M5 output comparisons differed for older ternary PQ2_0 at concurrency 4, including baseline repeats. That format does not use this kernel.
• Full CI and final reduced-patch M5 validation remain outstanding; this is for draft review.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. OpenAI Codex assisted with implementation, testing, benchmarking, analysis and this description. Contributor manual review and ownership are required before submission.

[bri-prism]

Tested on an M5 Pro (Metal), merged onto current prism (bda59ea); the merge is clean. test-backend-ops MUL_MAT PTQ1_0 passes 172/172 with GGML_METAL_PTQ1_MULTICOL off and on. Output with the switch on matches off on Ternary Bonsai 2 27B PTQ1_0 (llama-perplexity, ubatch 4: mean KLD ~0, max 3.7e-5, same top token 100%).

llama-bench, 27B PTQ1_0, -fa 1, two rounds, t/s: pp2 10.6 -> 37.6, pp3 15.0 -> 34.0, pp4 19.3 -> 39.9; pp1 and tg32 unchanged (25.1 / 25.8), and with the switch off it matches the prism base exactly. Without this, a two-column verify on Metal runs at under half the single-token speed, which is why speculative decoding (DFlash2 #261, MTP) loses on Metal today. Given there's no regression on the off path, it may be worth making this the default once it's been checked on one more Apple generation. LGTM.