llama.cpp b9434

TP: fix granularity for Qwen 3.5/3.6 + 3 GPUs (#23843)

• TP: fix granularity for Qwen 3.5/3.6 + 3 GPUs

• fix afmoe TP

UI:

• UI