Fix the OSS build error for fbgemm/torchrec (pytorch#1001)

Summary:
Pull Request resolved: pytorch#1001

OSS failure from D34840551 (pytorch@43c0f12):

- fbgemm_gpu: https://github.com/pytorch/FBGEMM/runs/5616872090?check_suite_focus=true
- TorchRec:  https://github.com/pytorch/torchrec/runs/5604628984?check_suite_focus=true

```
/home/ec2-user/actions-runner/_work/FBGEMM/FBGEMM/fbgemm_gpu/src/jagged_tensor_ops.cu:200:6: internal compiler error: in maybe_undo_parenthesized_ref, at cp/semantics.c:1739
   JAGGED_TENSOR_DISPATCH_DIMS();
```

~~Make num_jagged_dim to be only 1 / 2 before we can figure out a way to bypass such compiler issue.~~

**The issue**: we need to pass arg by value instead of by reference for some compiler version with macro.

Reviewed By: colin2328, jasonjk-park, brad-mengchi

Differential Revision: D35028747

fbshipit-source-id: 04da934c5dbdf5eafaf9dce8a081bf7f01344b12

fbgemm_gpu/include/fbgemm_gpu/sparse_ops_utils.h

@@ -264,8 +264,11 @@ constexpr uint32_t cuda_calc_block_count(
 }
 
 // Used in jagged_tensor_ops.cu and jagged_tensor_ops_cpu.cpp
+// Passing lambda exp argument by value instead of by reference to avoid
+// "internal compiler error: in maybe_undo_parenthesized_ref" error for specific
+// compiler version.
 #define JAGGED_TENSOR_DISPATCH_DIMS()                                         \
-  AT_DISPATCH_INDEX_TYPES(x_offsets[0].scalar_type(), "jagged_indices", [&] { \
+  AT_DISPATCH_INDEX_TYPES(x_offsets[0].scalar_type(), "jagged_indices", [=] { \
     switch (num_jagged_dim) {                                                 \
       case 1:                                                                 \
         INVOKE_KERNEL_WITH_DIM(1);                                            \