[torch api] Update external tensor handling for large attributes

[justinchuby]

          I am fine with adjusting the implementation later on, as long as we are fine with the API here. I think we may be better off making all tensors (initializers and node-attributes) of size above a cutoff limit external, leaving the rest internal. But I am fine if the cutoff limit is hard-coded in this logic for now ... only if we want to expose this to users, does it get a bit more complicated.

Originally posted by @gramalingam in #1832 (comment)