Review and improve segmentation and classification of vector cubes

[gilbertocamara]

The current implementation of the space-first, time-later approach in sits uses the SLIC algorithm to segment the image. Then, the resulting polygons are sampled. The classification probabilities of the polygon are calculated as an average of the probabilities of the sampled time series. However, the results do not match those produced by time-first, space-later method. Also, the division of the image in chunks without a buffer creates artefacts. Therefore, we need to investigate possible improvements, as follows

1. Replace SLIC with SNIC and measure the improvements.
2. Investigate methods to set good default parameters to SLIC.
3. Before combining the probabilities, run a local Bayesian smoothing on the probabilities of the samples.
4. Test different methods for combining the probabilities of the samples inside each polygon.

Associated sits API function
sits_segment(), sits_slic, and sits_classify.vector_cube()