xdna : selective hybrid CPU/NPU execution for Qwen3.5

[maximbibik-itrex]

Summary

Adds an AMD XDNA (NPU) backend for ggml: weight MUL_MAT operations are
offloaded to the AMD XDNA2 NPU via IRON (mlir-aie), while all other ops
continue to run on the standard GGML CPU backend. Includes a selective
hybrid CPU/NPU execution path tuned for Qwen3.5.

This is a draft PR — opening it early to get feedback on the approach
and backend integration points before finalizing. Not ready to merge yet.

What's included

• New ggml-xdna backend (ggml/src/ggml-xdna/) with IRON-based kernel
  build pipeline (ggml/src/ggml-xdna/iron/)
• --device XDNA support for routing weight GEMMs to the NPU
• Selective hybrid execution: NPU handles GEMM, CPU handles the rest
• Documentation of the validated setup, dependencies, and known toolchain
  pitfalls in docs/backend/XDNA.md

Validated environment

Built and measured on one machine so far:

• ASUS ProArt PX13 HN7306EA, AMD Ryzen AI MAX+ 395 (Strix Halo), NPU aie2p
• Ubuntu 24.04.4, XRT 2.25.37, amdxdna driver 2.25.260102.56
• IRON: mlir-aie 1.3.4 + llvm-aie 21
• Models tested: Qwen3.5-0.8B (Q4_K_M, BF16), Qwen3.5-9B (Q4_K_M)

[github-actions]

Review Status

Current Status: ❌ PENDING
Approvals so far: none

Pending reviews: Needs 1 Management or Team Lead, and 1 more from Management, Team Lead, or Member.