Fix unit test and posteriors

[hyanwong]

Previously we claimed to return posteriors that were probabilities but the testing was broken, so we didn't notice that we weren't scaling the posteriors to add up to one (i.e. they weren't actually probabilities). This fixes the problem, but therefore changes the posteriors that are output. I think this is the right thing to do.

[hyanwong]

I coded it up so that the inside_outside function returns the actual posterior value from the method, rather than standardising the values into probabilities (that sum to one), as this might be useful for checking the algorithm, e.g. to see how much the total value changes if we don't normalise.

The conversion to probabilities is then done immediately after the posteriors are returned, in the get_dates() function.

[nspope]

LGTM! Only a small semantic quibble regarding "normalised" (I'd try to avoid using as an adjective in the docs when talking about anything that's rescaled but doesn't sum/integrate to 1), and a question regarding the overflow protection.

[nspope]

The use of "normalised probabilities" is a bit confusing here, because normalizing a probability distribution typically means scaling such that it sums/integrates to 1; not the sort of under/overflow-protection that is happening here

[hyanwong]

Thanks a lot. I agree about "normalisation" (I'm not sure I named that one). I think "standardised" would be better. I reckon I should change the name of the method to that.

[hyanwong]

Merging. Will tackle renaming "normalise" to "standardize" in a followup PR. Thanks a lot for the feedback @nspope