hexagon: handle multi-sequence in concat_2d

[jhen0409]

Overview

Currently the op_concat in ggml-hexagon will fallback from 2d transposed to generic path when n_seqs greater than 1, especially in qwen35.cpp, which it will be slower than the fast path on PP. (See the CONCAT numbers in Additional information)

Since each plane is an independent 2D transposed concat, we should be able to iterate them instead of falling back to generic.

Additional information

Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4:

Model	npl	test	master (t/s)	patch (t/s)	change
Qwen3.5-4B Q4_0	1	pp512	373.07	376.00	+0.8%
Qwen3.5-4B Q4_0	1	tg32	10.62	10.59	-0.3%
Qwen3.5-4B Q4_0	2	pp512 x 2	49.36	181.24	+267.2%
Qwen3.5-4B Q4_0	2	tg32 x 2	13.49	14.11	+4.6%
Qwen3.5-4B Q4_0	4	pp512 x 4	49.86	176.66	+254.3%
Qwen3.5-4B Q4_0	4	tg32 x 4	14.59	15.47	+6.1%
granite-4.0-h-micro Q4_0	1	pp512	243.14	244.61	+0.6%
granite-4.0-h-micro Q4_0	1	tg32	8.73	8.71	-0.2%
granite-4.0-h-micro Q4_0	2	pp512 x 2	155.97	226.01	+44.9%
granite-4.0-h-micro Q4_0	2	tg32 x 2	10.75	10.99	+2.2%
granite-4.0-h-micro Q4_0	4	pp512 x 4	153.17	221.45	+44.6%
granite-4.0-h-micro Q4_0	4	tg32 x 4	13.52	14.04	+3.9%

Per-call CONCAT time (GGML_HEXAGON_PROFILE=1, one run per npl)

Model	npl	CONCAT shape (src0 x src1 -> dst)	calls	master (ms)	patch (ms)	change
Qwen3.5-4B Q4_0	1	3:8192 x 256:8192 -> 259:8192	48	1.265	1.269	+0.3%
Qwen3.5-4B Q4_0	1	3:8192 x 1:8192 -> 4:8192	768	0.233	0.233	+0.2%
Qwen3.5-4B Q4_0	2	3:8192:2:1 x 128:8192:2:1 -> 131:8192:2:1	96	138.243	1.356	-99.0%
Qwen3.5-4B Q4_0	2	3:8192:2:1 x 1:8192:2:1 -> 4:8192:2:1	768	0.752	0.469	-37.6%
Qwen3.5-4B Q4_0	4	3:8192:4:1 x 64:8192:4:1 -> 67:8192:4:1	192	154.644	1.635	-98.9%
Qwen3.5-4B Q4_0	4	3:8192:4:1 x 1:8192:4:1 -> 4:8192:4:1	768	1.500	0.927	-38.2%
granite-4.0-h-micro Q4_0	1	3:4352 x 256:4352 -> 259:4352	72	0.646	0.649	+0.4%
granite-4.0-h-micro Q4_0	1	3:4352 x 1:4352 -> 4:4352	1152	0.130	0.130	+0.3%
granite-4.0-h-micro Q4_0	2	3:4352:2:1 x 128:4352:2:1 -> 131:4352:2:1	144	15.102	0.696	-95.4%
granite-4.0-h-micro Q4_0	2	3:4352:2:1 x 1:4352:2:1 -> 4:4352:2:1	1152	0.395	0.252	-36.3%
granite-4.0-h-micro Q4_0	4	3:4354352:4:1	288	15.049	0.811	-94.6%
granite-4.0-h-micro Q4_0	4	3:43552:4:1	1152	0.783	0.494	-36.8%

The remaining gap I found for qwen35 so far is ssm_out MUL_MAT. In #29197, it uses MATMUL_4D_REPACKED_IMPL instead of falling back to CPU. In my experiments, flatten to 2D and use HMX could be more faster.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, I use Claude Code to track the issue, and did validation and review.

[max-krasnyansky]

@lhez please review/ack