Remove metadata on delegate getitem nodes (#342)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/342

mcr229 filed an issue which is that the delegate's getitem nodes (the
nodes pointing to each result of the call_delegate call) contain the metadata
of the original nodes, specifically the source_fn metadata. This causes an
issue because if we have 2 calls to to_backend, the first call will partition
torch.nn.Linear using source_fn metadata, and create a call_delegate node along
with getitem calls which now contain the torch.nn.Linear source_fn metadata.
When a second to_backend call comes along, if it also wants to partition based
on torch.nn.Linear source_fn metadata, it will incorrectly partition the
getitem nodes to the delegates made by the first to_backend call.

Implementation wise, this happens because the fuse_as_graphmodule function will
automatically propagate metadata of the nodes being partitioned, to the getitem
nodes. So, we will need to insert an extra pass to remove the metadata on these
nodes. Note that this will also remove the "val" metadata, but we will bring it
back in final the ExportPass() call at the end of to_backend.

Reviewed By: digantdesai, cccclai

Differential Revision: D49264387

fbshipit-source-id: 4755b2561f926738524c421d370525611fbfe4c9