- refactor(mge): add support for optimizer.step().clear_grad() idiom

- refactor(mge): rename some methods of GradManager

- refactor(mge): remove tensor_nn and TensorDict

- refactor(mge): remove Buffer

- refactor(mge): remove requires_grad flag

- refactor(mge): add a default grad=None attribute to Tensor

- refactor(mge): deprecation for 1.0

GitOrigin-RevId: 3b723d938747e7a7d765ba3716b2e08f5223f62a