llama.cpp PR 26291 speeds up RPC model loading from 5min to 1:30

PR 26291 cuts llama.cpp RPC model loading from 4min54sec to 1min38sec (~300% faster) using the new GGML_RPC_LOAD_THREADS=12 variable, bench-marked with a 300GB model on RTX 4060 Ti systems with DDR4 and DDR5 memory.