perf(llama-cpp-rocm): tune for Strix Halo (gfx1151)

- Restrict rocmGpuTargets to gfx1151 (Radeon 8060S, RDNA 3.5) — smaller
  closure, faster compile, no wasted device kernels.
- Enable GGML_HIP_ROCWMMA_FATTN: rocWMMA-backed flash attention is a
  major win on RDNA3+ for the GPU-offloaded attention path.
- Enable GGML_HIP_GRAPHS to lower per-token launch overhead.
- Add rocwmma to buildInputs to satisfy the WMMA path.

llama-server on halo runs with -ngl 99 --flash-attn on, so these flags
target the hot path. CPU-side AVX-512 was skipped intentionally — Zen 5
has it, but with full GPU offload the CPU paths barely run.

overlays/unstable/default.nix

@@ -12,7 +12,13 @@ final: prev: {
     # nodejs_20
     ;
 
-  llama-cpp-rocm = channels.unstable.llama-cpp-rocm.overrideAttrs (_: {
+  # Tuned for Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S, gfx1151).
+  llama-cpp-rocm =
+    (channels.unstable.llama-cpp.override {
+      rocmSupport = true;
+      rocmGpuTargets = [ "gfx1151" ];
+    }).overrideAttrs
+      (prevAttrs: {
         src = prev.fetchFromGitHub {
           owner = "am17an";
           repo = "llama.cpp";
@@ -23,6 +29,15 @@ final: prev: {
           '';
         };
         npmDepsHash = "sha256-k62LIbyY2DXvs7XXbX0lNPiYxuYzeJUyQtS4eA+68f8=";
+
+        buildInputs = (prevAttrs.buildInputs or [ ]) ++ [
+          channels.unstable.rocmPackages.rocwmma
+        ];
+
+        cmakeFlags = (prevAttrs.cmakeFlags or [ ]) ++ [
+          "-DGGML_HIP_ROCWMMA_FATTN=ON"
+          "-DGGML_HIP_GRAPHS=ON"
+        ];
       });
 
   /*