### What this PR does / why we need it? Past: npu_moe_gating_top_k can only support 'group_count=256' pattern Now: 1、npu_moe_gating_top_k support all size of group_count 2、the functionality of `torch_npu.npu_moe_gating_top_k_softmax` are included in `torch_npu.npu_moe_gating_top_k` CANN: depends on 8.3.RC1 Performance: 1. GLM4.5-w8a8, TPS improve 6% 2. Qwen3, the same as before - vLLM version: v0.11.0 - vLLM main: https://github.com/vllm-project/vllm/commit/2918c1b49c88c29783c86f78d2c4221cb9622379 Signed-off-by: 1092626063 <1092626063@qq.com>