llama.cpp trims CUDA MoE overhead

This looks like the right kind of low-level optimization: small algorithmic and arithmetic changes that barely move dense cases, but unlock measurable gains where CUDA bookkeeping was the bottleneck.

• –The main win comes from reducing integer-division overhead in MMQ and deciding between stream-k and tiling based on estimated efficiency loss, not a hard-coded MoE rule.
• –Benchmarks show the uplift is concentrated in MoE prompt processing, especially on multi-GPU setups; dense-model runs are mostly flat, which suggests the patch is targeted rather than risky.
• –The implementation is cautious: `kbc` moved to 32-bit with host-side overflow assertions, while the remaining critical arithmetic stays 64-bit where needed.
• –For llama.cpp users running MoE models on NVIDIA hardware, this is the kind of change that matters more than a flashy headline feature because it improves the hot path directly.