[Dynamo] Refine CPU fallback for TD+XLA (#5000)
* Support CPU fallback for unsupported ops in dynamo
* Update imports
* move partition outside of extract compile graph
* attempt to return compiled graph instead
* replace submodule with a function call instead
* make materialization check return true when input XLA tensors do not have XLAData or IR
* fixed op support check
* fix lint issue
* add tests
* fix cpp lint
* modify tests
* Include module into fallback checking
* Add more tests with numerical checks
* Rename test file
* Fix metrics in test_dynamo.py
* Run linter
* Update dynamo fallback logic to support in-place operations
* Fix fault tolerance to fix GPU CI
* Clear pending IRs after initial fetch of fallback ops
* Change the metric asserts back to original
* Revert irrelevant comments in test_dynamo.py
* Refactor fallback dynamo tests
* Run linter
* add call_method to fallback check list
* Update fallback unit tests
---------
Co-authored-by: Ubuntu <ubuntu@ip-172-31-13-117.us-west-2.compute.internal>
Co-authored-by: seanlatias <seanlatias@gmail.com>
Co-authored-by: Wonjoo Lee <wonjoo@example.com>
