This is only a suggestion, but one thing that is possible as an alternative to deliberate smoothing is introduction of additional (fake) data points through which the curve should pass.

The condition to insert such points is to detect if there is a sharp reversal in direction between any two consecutive (real) points. This could be done by comparing the angle between directions of vectors containing any two consecutive nodes and if this angle is lower than certain threshold, the point is considered a 'reversal' point.

Once such reversal points are identified you can introduce P (fake) points, where the integer P depends on how smooth you would like these transitions to be. E.g. if you don't mind a U-shape, then you could introduce just one fake point per each reversal point that is slightly offset in the direction orthogonal to the reversal direction.

One known method is to interpolate points using Dubins paths. It provides the shortest curve between any two curves, while satisfying a minimum curvature. One implementation is described in this paper, called Markov-Dubins interpolation.

However, I do not recommend this approach because the curves are not smooth (there are discontinuities in the acceleration due to fixed levels of curvature).

Not at the moment, at least not in scipy. Sadly.