Add unit tests for the MoE WNA16 Marlin GEMM op.
Source: custom_ops/gpu_ops/moe/moe_wna16_marlin_gemm.cu
Registration: custom_ops/gpu_ops/cpp_extensions.cc
Test file: tests/operators/test_moe_wna16_marlin_gemm.py
This is a quantized MoE GEMM kernel. Tests should compare output against a dequantized FP16 reference. Cover different expert counts, hidden sizes, and token counts.
Branch: task/033-moe-wna16-marlin-test