[composite compliance] preserve stride correctly for non-contiguous tensor with requires_grad=True (#81035)
Fixes https://github.com/pytorch/pytorch/issues/80858
Also removes the skips on `ravel`, `lstsq` and `lstsq_grad_oriented`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81035
Approved by: https://github.com/zou3519