Use release/acquire ordering for GatherND error flag

Address review feedback: with relaxed ordering on separate atomics,
the acquire of has_invalid_index could observe true while err_index
is still at its default. Store err_index first (relaxed), then
publish has_invalid_index with release. Load the flag with acquire
to guarantee err_index visibility.

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

Author: bmehta001

onnxruntime/core/providers/cpu/tensor/gather_nd.cc

@@ -91,7 +91,7 @@ Status GatherNDBase::PrepareForCompute(const TensorShape& input_shape, const Ten
   // err_index == 0 to mean "no error" creates a collision when a zero-sized
   // dimension makes index 0 invalid — the error goes undetected.
   std::atomic<bool> has_invalid_index{false};
-  int64_t err_index = 0;
+  std::atomic<int64_t> err_index{0};
   p.element_bytes = bytes_per_value;
   p.element_count_per_slice = slice_size;
   p.bytes_per_slice = p.element_bytes * p.element_count_per_slice;
@@ -110,8 +110,10 @@ Status GatherNDBase::PrepareForCompute(const TensorShape& input_shape, const Ten
       const auto upper_limit = input_shape[SafeInt<size_t>(batch_dims_) + dim_idx];
       const auto lower_limit = -upper_limit;
       if (index < lower_limit || index >= upper_limit) {
-        has_invalid_index.store(true, std::memory_order_relaxed);
-        err_index = index;
+        // Store err_index first, then publish via has_invalid_index with release
+        // so the acquire load on the flag guarantees err_index is visible.
+        err_index.store(index, std::memory_order_relaxed);
+        has_invalid_index.store(true, std::memory_order_release);
         break;
       }
       if (index < 0) index += upper_limit;
@@ -130,9 +132,10 @@ Status GatherNDBase::PrepareForCompute(const TensorShape& input_shape, const Ten
         }
       });
 
-  return !has_invalid_index.load(std::memory_order_relaxed)
+  return !has_invalid_index.load(std::memory_order_acquire)
              ? Status::OK()
-             : ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, "invalid index found, index = ", err_index);
+             : ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, "invalid index found, index = ",
+                              err_index.load(std::memory_order_relaxed));
 }
 
 template Status GatherNDBase::PrepareForCompute<int32_t>(const TensorShape&,