KSD metric using precision threshold

[gw265981]

What's the issue?

The MMD metric has been changed from using a precision threshold to just setting the floor to 0 in order to deal with precision issues (#862).

The KSD metric is for now unchanged because: 1) it hasn't been observed to suffer from the same issue in practice; 2) it could theoretically be negative if $p(x)$ and/or $\lambda$ are very large.

If we are confident that the KSD metric should always be non-negative or if we encounter similar precision issues, we can enact the same change for KSD in the future.