derivative for GELU

[KGrewal1]

Implementation of backprop derivative for GELU using the analytical derivative of the GELU approximation
$$0.5v\left(1+\tanh\left(\sqrt{\frac{2}{\pi}} v(1+0.044715v^{2})\right)\right)$$
as
$$0.5 + (0.398942 v + 0.0535161 v^3) (1 - \tanh^{2}(0.797885 v + 0.0356774 v^{3})) + 0.5 \tanh(0.797885 v + 0.0356774 v^{3})$$

[LaurentMazare]

Looks nice, could you also add a test to unary_grad and ensure that it matches the PyTorch output?
Also probably better to compute arg.powf(3.) only once rather than twice.

[KGrewal1]

Tests added (using same input tensor as other tests) and changed to cube tensor only once

[LaurentMazare]

Great, thanks for the PR!