ggml: optimize concat op by replacing per-element memcpy with row-level memcpy

[sirohikartik]

Overview

Optimize ggml_compute_forward_concat_any by replacing per-element memcpy with row-level memcpy.

The original implementation called memcpy once per scalar element with a branch inside the innermost loop to select between src0 and src1. For a typical KV-cache concat shape [4096 x 1 x 16 x 1] along dim=2 this results in 65,536 separate memcpy calls of 4 bytes each.

This PR splits the loop into two separate regions (one per source tensor) eliminating the per-element branch, and collapses the i0 loop entirely to copy one full row per memcpy call instead of one element.

Benchmark

Isolated microbenchmark using identical tensor layout and loop logic.
Shape: [4096 x 1 x 16 x 1], concat dim=2, fp32, 200 runs
Apple Silicon (M1 Air):

	old	new	speedup
warm cache	1.61 ms/call	0.012 ms/call	133x
cold cache	1.59 ms/call	0.019 ms/call	83x

Cold cache measured by flushing 64MB through memory before every call.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES I used Claude Sonnet 4.6 to help understand the existing
  code, identify the inefficiency, and verify the approach. All changes reviewed and benchmarked by me.

[sirohikartik]

Hi @ggerganov I think this is ready for review.