metal : PTQ1_0 multi-column mat-vec for 5-8 columns

[jasontitus]

Overview

#262 added GGML_METAL_PTQ1_MULTICOL=1, a PTQ1_0 mat-vec for 2-4 columns; 5-8 columns still take mul_mv_ext. Batched decode, MTP verification and multi-request steps often produce 5-8 columns (4 requests with MTP draft 1 are 8).

This PR splits 5-8 columns into two tiles of at most four (5 -> 3+2, 6 -> 3+3, 7 -> 4+3, 8 -> 4+4) on the existing c2/c3/c4 kernels:

• a partial last tile clamps its column reads to the last valid column and skips the stores past it
• a bool function constant (FC_MUL_MV + 5) removes those checks when every tile is complete, so 2-4 columns compile to exactly the metal : add PTQ1_0 mat-vec for two to four columns #262 kernel
• GGML_METAL_PTQ1_MULTICOL_MAX (clamped to 4..8, default 8) caps the widths the kernel takes; 4 gives the metal : add PTQ1_0 mat-vec for two to four columns #262 behavior

Additional information

M5 Max, Ternary Bonsai 2 27B PTQ1_0, this branch, GGML_METAL_PTQ1_MULTICOL=1 in both arms, max 4 (#262) vs max 8, three A-B-B-A quartets with 8 s cooldowns (tok/s; MTP rows are llama-server with draft-mtp, draft 1, aggregate, 128 greedy tokens):

Case	Columns	#262	This PR	Speedup [quartet range]
pp4 (control)	4	69.86	70.34	1.007 [1.000-1.013]
pp5	5	45.14	61.82	1.370 [1.368-1.373]
pp6	6	30.76	74.78	2.433 [2.373-2.552]
pp7	7	34.91	73.20	2.097 [2.059-2.164]
pp8	8	40.45	86.82	2.146 [2.138-2.161]
tg128 (control)	1	43.67	43.39	0.994 [0.987-1.003]
MTP, 3 requests	6	24.48	52.23	2.134 [2.123-2.148]
MTP, 4 requests	8	28.53	55.07	1.932 [1.888-2.023]

Generated tokens identical in every pair.

Correctness:

• test-backend-ops -o MUL_MAT: 1493/1493 with the switch off, on with max 4, on with max 8 (the partial-tile pipelines compile and run), and with an out-of-range max. New cases: 5-7 columns with row tails and broadcast, and 5 and 7 with a strided B.
• Full model, batches of 1-8 tokens (108 positions): switch off, logits bitwise equal to upstream; max 4, bitwise equal to metal : add PTQ1_0 mat-vec for two to four columns #262; max 8 first differs at the first 5-token batch (different summation order than mul_mv_ext), max relative difference 3.7e-5, 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.