llama.cpp b9095 adds NCCL-free tensor parallelism

This is a meaningful infrastructure step for local inference: it lowers the dependency burden for multi-GPU tensor parallelism and makes dual consumer Blackwell rigs easier to bring up.

• –The new internal AllReduce is explicitly NCCL-free, which matters most on desktop-class NVIDIA setups where NCCL can be a setup friction point
• –The implementation is still narrow in scope, so this is a practical win for specific dual-GPU workflows rather than a universal multi-GPU answer
• –The release notes say the kernel works on Volta-or-newer NVIDIA GPUs, so the impact is broader than the Reddit title implies
• –`GGML_CUDA_ALLREDUCE` and `--allreduce` make it easy to compare internal vs NCCL paths and debug regressions
• –For local model builders, this kind of plumbing change can improve throughput and reliability without changing the model stack