ggml 1.0.0 adds FP8 support and breaks compatibility with older RPC peers

ggml 1.0.0 adds FP8 support on CUDA and new execution paths across hardware backends, according to its release notes. The release also removes deprecated public interfaces and legacy masking operations, updating its remote procedure call protocol to reject incompatible peers. The scheduler now avoids uploading inputs that computations do not read. A cache for mixture-of-experts models can compute a share of cache misses on the CPU and copy results back asynchronously. Other additions include an Intel FlashAttention prefill kernel on Vulkan and a device-side mixture-of-experts scheduler on SYCL. The RPC protocol moves to v9, with faster loading and graceful failure on unreachable endpoints or failed handshakes.