fix: clamp to sphere in finite diff computation

[alvaro-budria]

This is a PR for fixing issue #85. In the gradient computation by finite differences, the input points were being clamped to [0, 1] range, rather than [-radius, radius].

[bennyguo]

Thanks!