add ROCm backend support

[alantsev]

add ROCm backend support

This PR adds ROCm support as a third GPU backend alongside the existing CUDA and Metal backends.

• Makefile — Added GPU_BACKEND=rocm with hipcc, configurable via ROCM_PATH and ROCM_ARCH. Existing CUDA and Metal builds are unchanged.
• ds4_cuda.cu — Now builds for both CUDA and ROCm from a single source via #ifdef __HIP_PLATFORM_AMD__, using FULL_WARP_MASK/MASK_T to handle the warp width difference.
• ds4_rocm.h (new) — Compatibility header mapping CUDA runtime / cuBLAS to HIP equivalents, plus AMD implementations of __vcmpne4, __vsub4, __dp4a.

Build

make GPU_BACKEND=rocm
make GPU_BACKEND=rocm ROCM_ARCH=gfx942

Benchmarks — AMD Radeon 8060S (gfx1151), ROCm 6.x:

ctx_tokens | prefill_tps | gen_tps
     2,048 |       82.36 |    7.40
     8,192 |       80.59 |    7.03
    16,384 |       80.51 |    6.69
    32,768 |       79.62 |    6.10
    49,152 |       78.77 |    5.59
    65,536 |       78.04 |    5.24

[janimo]

@alantsev nice, tested on Ubuntu 26.04/Strix Halo with both the distro's ROCm 7.1 and AMD's 7.2.3 packages.
I am getting slightly lower prefill numbers (~70) than you in the bench.

[ucjonathan]

Lot of these $2-3k machines out there. My friend has one. Glad you guys are doing the work to support this platform as well. Impressed that the community is going from Apple Metal -> CUDA -> ROCm support in just about a week. Someone is going to need to get a Strix machine in @antirez hands though before it will be merged in. I really think there needs to be some corporate level sponsorship on the hardware side for this project between Apple / Nvidia / AMD / Intel etc.