🐛 [Bug] LongType introduced in Conversion process by prim::NumToTensor

[bowang007]

Bug Description

We have a graph like this:

  %14 : int = aten::size(%t.1, %15) 
  %16 : Tensor = prim::NumToTensor(%14) 
  %17 : Tensor[] = prim::ListConstruct(%13, %16)
  %image_size.1 : Tensor = aten::stack(%17, %3) 

This fails with an error:

terminate called after throwing an instance of 'c10::bad_optional_access'
  what():  bad optional access

This happens because we get an int from aten::size, then in prim::NumtoTensor we are using this function https://github.com/pytorch/pytorch/blob/8a7c9a5e01a24d465126210234aa9d3775b25032/aten/src/ATen/ScalarOps.h#L29, which will introduce a long tensor that incur the error.

Expected behavior

Truncate the long type here, should be fixed soon.