Use `index_select` instead of fancy indexing in `batched_mm_experts_forward`
Replace tensor[index] fancy indexing with torch.index_select() for expert
weight selection. Both produce identical results, but index_select is more
explicit and improves compatibility with compiler backends that struggle
with vector-indexed gather operations (e.g., Neuron's NCC_IDLO901).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>