Make Clustering more intuitive

[Muennighoff]

Some great ideas from @hanxiao in this tweet https://x.com/hxiao/status/1818352657246896226 mainly for the scatterplot, always use fixed & squared xlim and ylim; or overplot a unit sphere as the reference; or rotate the two scatter plots so that their principal vectors are allgned

If someone has bandwidth to play with this a bit and put some proposals in this issue that'd be amazing! Code-wise should be pretty simple, just modify models.py where the clustering is done :)