[inductor] Do not try to shape-pad symbolic-sized tensors (#100738)
Summary: We use benchmarking to decide whether to pad tensors for mm alignment, but if the sizes are symbolic, we can't really do that.
Test Plan:
```
pytest test_torchinductor_dynamic_shapes.py -k padding
```
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100738
Approved by: https://github.com/jiawenliu64, https://github.com/ngimel