investigate kohonen neural net approach for quantization

[dankamongmen]

libsixel currently uses a fast median cut approach to quantization, borrowing significant code (iirc) from netpbm. Multiple other quantization methods exist, including self-organizing kohonen nets, octrees, and K-means clustering. of these, kohonen nets seem to me to be the most intriguing. KNNs (not to be confused with k-nearest-neighbors, erp) are used in neoquant, and are also the strategy employed in the rust color-quantize crate.

write up an implementation, and benchmark it against the current one for both accuracy and performance. if it's promising, investigate replacing the median cut approach.