[Fix] : Stabilize Backpropagation for Sum and Mean

[Advaitgaur004]

This PR tackles an important bug in ctensor, making backpropagation through sum and mean operations correct.

What Was Happening?

Previously, calling Tensor_backward() on a tensor that was the result of a Tensor_sum or Tensor_mean (with a dim argument) would crash. This was because the upstream gradient had a "squashed" shape (e.g., {2}) while the local gradient had the original, larger shape (e.g., {2, 3}), making them incompatible for the chain rule's multiplication step.

Fix:

• The Tensor_backward function now dynamically compares the input and output shapes of a reduction operation to deduce which dimension was removed.
• The "Unsqueeze" Trick: Once the reduced dimension is found, I "unsqueeze" the upstream gradient (e.g., from {2} to {2, 1}), making it ready for broadcasting.
• Fixing the mean Derivative: I also found and fixed a bug in GradFn_mean where it was incorrectly calculating the gradient value based on the output tensor's size instead of the input's. It now correctly computes the 1/N gradient.

As a result, the sum and mean tests, including chained operations with nn_linear and nn_relu, are now all passing.