[vLLM]Skip Marlin MoE tests

[sarah12121212]

Update the vLLM patch file to skip CUDA/NVIDIA-specific Marlin moe kernels, like ROCm currently does.

Changes

• Adds XPU to the existing current_platform.is_rocm() skip guards for:
  • test_fused_marlin_moe
  • test_fused_marlin_moe_with_bias
  • test_fused_marlin_moe_non_gated
  • test_batched_fused_marlin_moe
• Removes the above tests from the default and xe2 moe skiplists.

CI Validation:

PVC:
BMG: