better modeling

[lukeshimanuki]

currently it uses independent single variable gaussians, meaning the PDFs will always be aligned with the axes
this is problematic because usually it sees things in a direction not aligned with the axes, so we end up using the variance associated with the most variant axis in our observations
you should instead use multivariate gaussians
you'll probably want to approximate the resulting PDF as a gaussian when adding evidence because representing arbitrary combinations makes the math very annoying

don't try this unless you actually understand how it works and you have a ton of time to work out the bugs