ShapeEnv.tracked_fakes can hold CUDA tensors after compile. this clears it and removes the leak-check skip from the currenet regression

Testing:
- PYTORCH_TEST_CUDA_MEM_LEAK_CHECK=1 python test/dynamo/test_misc.py MiscTestsDeviceCUDA.test_interpolate_propagate_real_tensors_cuda
  - passed

Closes #190093

Pull Request resolved: https://github.com/pytorch/pytorch/pull/193015
Approved by: https://github.com/jansel