hexagon: use DMA for contiguous dim1 CONCAT

[kurquhar]

Overview

Use DMA for contiguous dim1 CONCAT on hexagon backend

Additional information

Motivated by latency analysis of LFM2.5-DSpark, which showed that CONCAT occupied ~25% of the profiled HTP time (excluding CPU time).

The kernel implementation was leveraged from two other kernel classes: the scalar concat-ops.c and DMA-based DDR->DDR in cpy-ops.c.

The kernel is gated to contiguous tensors of the same type (FP32, FP16, or INT32), concat along dim1, and dim0/2/3 same size.

Eval

Compared to the baseline [ec92815], this change improved LFM2.5-DSpark pipeline performance by over 23% during ABBA testing. Several portions of this pipeline run on CPU by design, so thermal throttling can impact measurements significantly:

• Baseline: mean 30.46 tok/s
• New: mean 37.65 tok/s
• Gain: +7.19 tok/s, +23.6%
• All endpoints matched: 176/692 accepted/drafted, mean length 3.23.
• Initial 10min cooldown and additional 180s between phases

Command used for eval:

python scripts/snapdragon/run.py --install-dir <pkg-location> --devices HTP0 -- \
  llama-cli \
  -m <model dir>/LFM2.5-2.6B-Q8_0.gguf \
  -md <model dir>/LFM2.5-2.6B-DSpark-Q8_0.gguf \
  --spec-type draft-dspark --spec-draft-n-max 10 --spec-draft-n-min 0 \
  --no-spec-draft-backend-sampling \
  -c 4096 -ngl 99 -ngld 99 -fa on -n 256 \
  -p "The sky appears blue because" \
  --temp 0 --no-warmup --no-jinja --reasoning off --no-display-prompt -st -lv 4

Backend Op Tests

Compared to the baseline [ec92815], this change does not change op test results.

• Baseline: 48 [OK]
• New: 48 [OK]

Command used for op tests:

python scripts/snapdragon/run.py --install-dir <pkg-location> --devices HTP0 -- \
  test-backend-ops -o CONCAT -b HTP0

Requirements

• I have read and agree with the contributing guidelines yes
• AI usage disclosure: Yes, OpenCode+Terra was used for reasoning, coding, and A/B testing. All generated code has been reviewed by me (human).

Assisted-by: OpenCode

[max-krasnyansky]

@lhez please review/ack when you get the chance