add dispatch_ffn_combine_bf16 kernel for deepep

[zuje123]

DeepEp supports dispatch_ffn_combine_bf16, which designed to enhance the performance of certain deep learning models by fusing operations and without quant.

Based on dispatch_ffn_combine_bf16, the following adaptations were made:

1. Changed the parameter types of weight1, weight2, scale1, and scale2 from TensorList to Tensor.
2. Set cce-auto-sync=off and resolved synchronization issues.
3. Fixed the issue where moe_init_routing_v2 encountered the error "VEC supports illegal configurations in commands" when bs=0.
4. Add validation for HCCL_BUFFSIZE.

Test with Qwen3-30B-A3B:

[gemini-code-assist]

Warning

Gemini is experiencing higher than usual traffic and was unable to create the review. Please try again in a few hours by commenting /gemini review.