Triton input-staging kernel for DeepSeek V4 MegaMoE.
Quantizes hidden states to fp8 with E8M0 group scales and repacks the routing top-k tensors into the int64/float32 layout that the DeepGEMM MegaMoE kernels consume.