Add gfx950 target for ROCm 7.2

[superm1]

AMD's binary builds do include GFX950.

CC @IMbackK , @slojosic-amd, @hammmmy

[superm1]

Here is test run in my fork: https://github.com/superm1/llama.cpp/actions/runs/22309686343/job/64538059208

[IMbackK]

gfx950 is not supported.

it will match

llama.cpp/ggml/src/ggml-cuda/common.cuh

Line 90 in 9051663

#define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)

but not

llama.cpp/ggml/src/ggml-cuda/vendors/hip.h

Line 198 in 9051663

#if defined(CDNA3) || defined(CDNA2) || defined(CDNA1)

which will cause the host and device code to get out of sync.

Unfortunately the hip compiler dose not provide a better solution to detect it on the device side at compile time other than to add every device family.

Someone with access to gfx950 hardware will have to add it, then test, and ideally also tune the parameters these macros are for first.

[superm1]

@slojosic-amd can you help?

[CISC]

Since this is not built by any other CI I'd like to see a fresh successful run please.