Get `required_grad` check before the gradient calculation

[tomiock]

Currently, the required grad happens in the gradient accumulation function, after the actual gradient calculation.

Changing this would saved some operations. For example when having data (that do not require a gradient), the gradient for all the datapoints is going to get calculated. This can be very expensive, especially for high dim data.