Use fx.replace_pattern for removing empty_like+fill in nvFuser+PrimTorch execution (#89132)
I learned about `torch.fx.replace_pattern` and it's a cleaner way of removing unnecessary tensor materialization from the graph coming from tracing C++ code `1 - tensor`.
Test:
```
python -m pytest test/test_prims.py -k "test_silu_backward_no_filled_tensor"
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89132
Approved by: https://github.com/mruberry, https://github.com/jjsjann123
