improve retain_grad to have same functionality as in torch

[aanurraj]

Description

Currently, we have an issue if we try to do an operation on two mpc tensors one with requires grad = True and the other with requires grad = False, and run backward to calculate the gradient.

result of the two of the above-mentioned tensors gives grad which is not the case if we do a similar operation on the torch.

Are you interested in working on this improvement yourself?

• Yes, I am.

Additional Context

• complete the mentioned task.
• complete the todo to add the test for the same at test_backward_with_one_requires_grad in mpc_tensor_tests.py