ggml-cuda: fix UMA memory detection for HIP/ROCm on AMD APUs

AMD APUs report prop.integrated=1 which triggers the UMA memory
path from #17368. This overrides hipMemGetInfo() (accurate) with
/proc/meminfo MemAvailable (too low), losing ~30 GiB on a 128GB
Strix Halo system.

For HIP builds, only enter the UMA path when GGML_CUDA_ENABLE_UNIFIED_MEMORY
is explicitly set. This preserves correct behavior for both cases:
- Default: hipMemGetInfo() reports accurate TTM-backed memory
- GGML_CUDA_ENABLE_UNIFIED_MEMORY=1: /proc/meminfo is used (system RAM mode)

Tested on AMD Ryzen AI MAX+ 395, Radeon 8060S (gfx1151), 128GB, ROCm 7.1.

Fixes: ggml-org#18159

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -4569,7 +4569,14 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
 
     // Check if UMA is explicitly enabled via environment variable
     bool uma_env = getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr;
+
+#if defined(GGML_USE_HIP)
+    // On AMD APUs, prop.integrated is true but hipMemGetInfo() already returns
+    // the correct TTM-backed memory. Only use the UMA path when explicitly requested.
+    bool is_uma = uma_env;
+#else
     bool is_uma = prop.integrated > 0 || uma_env;
+#endif
 
     if (is_uma) {
         // For UMA systems (like DGX Spark), use system memory info