RocmAadEngine

ROCm/HIP replay engine. Priority sits above SIMD and below the Vulkan engine, so on this repo's AMD-APU dev box Nabla.model(...).fastest() keeps choosing Vulkan (the RADV compute path has proven the more reliable of the two here); ROCm is selected explicitly with -Dnablatensor.engine=rocm or .on("rocm"), which is the intended way to benchmark it.

Holds no reference to libamdhip64 beyond isAvailable(), which probes through the HipCompute facade, so merely enumerating engines cannot fail where ROCm is absent.