llama.cpp Adds RPC Tensor Parallelism

This is a major step toward practical multi-box local inference, bringing tensor-parallel execution to llama.cpp without a fork. The catch is that it remains highly dependent on fast interconnects and model support.

• –Adds asynchronous graph execution, cross-node all-reduce, graph caching, and 2D tensor transfers for RPC workers.
• –Pools memory across machines, making models possible that cannot fit on either box individually.
• –RDMA is the sweet spot; tests over gigabit networking showed tensor mode far behind layer splitting.
• –The feature is experimental, and users must keep RPC clients and servers on compatible builds.
• –llama.cpp documents tensor mode as parallelized but experimental, with flash attention typically required. [Multi-GPU documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs/multi-gpu.md)