[Test] Refactor `test_dynamic_shapes.py` (#194081)

- add hardware classification
- remove obsolete skip on XPU
- extract `test_flex_attention_foreign_fake_e2e` into separate class (device-dependent test)
- generalize `test_flex_attention_foreign_fake_e2e` by replacing hardcoded `"cuda"` with a generic device
- add `allow_xpu=True` to `TestTransferSymbolsFromForeignShapeEnvDevice`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/194081
Approved by: https://github.com/etaf, https://github.com/jansel