Z3: use .contiguous() for NCCL allgather send buffer; add comment

Signed-off-by: Abhishek <dalakotiashu150@gmail.com>

Author: adalakoti90

csrc/compile/z3.cpp

@@ -74,7 +74,8 @@ class Z3CustomOpExecutor : public CustomOpExecutor {
             const int64_t shard_elems = ds_tensor.numel();
 
             // Perform all-gather directly into the pre-allocated padded output buffer
-            ncclResult_t result = ncclAllGather(ds_tensor.flatten().data_ptr(),
+            // NCCL requires contiguous storage; use .contiguous() explicitly
+            ncclResult_t result = ncclAllGather(ds_tensor.contiguous().data_ptr(),
                                                 output_buf.data_ptr(),
                                                 shard_elems,
                                                 get_nccl_data_type(ds_tensor.scalar_type()),