metal : opt-in PQ2_0 mat-vec for two columns (GGML_METAL_PQ2_MULTICOL=1)

[jasontitus]

Overview

Two-column PQ2_0 products (two requests, or one request with an MTP draft of 1) take mul_mv_ext. This adds an opt-in two-column mat-vec, GGML_METAL_PQ2_MULTICOL=1, in the style of #262: each byte's three floors are computed once for both columns, and each column's collapse coefficients once for all rows. The collapse arithmetic is kernel_mul_mv_pq2_0_f32's. From three columns mul_mv_ext stays faster, so only two columns use it. Off by default; batched and broadcast src1 are handled like the other mat-vec kernels.

Additional information

M5 Max, Ternary Bonsai 2 27B PQ2_0, this branch, switch off vs on, three A-B-B-A quartets with 8 s cooldowns (tok/s; llama-server rates are aggregate, 128 greedy tokens):

Case	Off	On	Speedup [quartet range]
tg128 (control)	45.35	45.19	0.997 [0.987-1.006]
pp2	57.94	68.11	1.175 [1.172-1.181]
pp3 (control)	71.51	71.74	1.003 [0.995-1.011]
llama-server, 2 requests	48.12	55.51	1.154 [1.135-1.186]
llama-server, 1 request, MTP draft 1	45.00	52.42	1.165 [1.153-1.177]

Generated tokens identical in every pair.

Correctness:

• test-backend-ops -o MUL_MAT: all pass with the switch off and on; the new kernel runs for plain, batched (ne12 = 2) and broadcast (r2 = 2) cases. New cases: two columns with a row tail and broadcast, and with a strided B.
• Full model, batches of 1-8 tokens (108 positions): switch off, logits bitwise equal to upstream; switch on first differs at the first 2-token batch, max NMSE 3.9e-10, same top token at every position.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Claude (Anthropic) helped with the implementation, testing, measurements and this description; I reviewed the change and own it.