Summary:
Use `to_edge_transform_and_lower` in `export_llama` for XNNPack. As part of these changes, this also means that you cannot specify multiple backends in `export_llama` in the args, although I'm not sure if that is happening anywhere at the moment.

Closes https://github.com/pytorch/executorch/issues/8621

Performance regression benchmarking for xnnpack (on android) vs. past 3 days:
<img width="1427" alt="Screenshot 2025-02-24 at 11 39 52 AM" src="https://github.com/user-attachments/assets/1640cf2c-a579-491f-8940-7ccfbe464903" />

These benchmark numbers also normally fluctuate a bit across runs and these differences are within the usual fluctuation ranges.


Test Plan: See if CI passes

Differential Revision: D70124742

Pulled By: jackzhxng