Add clarifying comment for CUDA bounds guard per review

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: vraspar

onnxruntime/contrib_ops/cuda/bert/ngram_repeat_block_impl.cu

@@ -49,6 +49,8 @@ __global__ void banRepeatedTokens(const int64_t* __restrict__ tokens,
   if (is_banned == true) {
     auto token_to_be_banned = tokens_shm[col + no_repeat_ngram_size - 1];
     CUDA_KERNEL_ASSERT(token_to_be_banned >= 0 && token_to_be_banned < vocab_size);
+    // In release builds, silently skip OOB tokens rather than writing out of bounds.
+    // CUDA kernels cannot propagate Status errors to the host.
     if (token_to_be_banned >= 0 && token_to_be_banned < vocab_size) {
       lprobs[lprob_start + token_to_be_banned] = -std::numeric_limits<float>::infinity();
     }