Make Openxla dynamo backend take boxed input (#107260)
Fixes https://github.com/pytorch/xla/issues/5454
Also adding the inference(non-aot) backend back since we see a speed regression when using the aot-backend compared to the non-aot openxla backend. It is being tracked in https://github.com/pytorch/xla/issues/5430
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107260
Approved by: https://github.com/shunting314, https://github.com/jansel