Incorporate error into handicap adjustments

[dexonsmith]

Change get_handicap_adjustment to return a tuple with both rating and error, where the error comes from:

• ±0.5 for rank differences that aren't (almost) the same as the number of handicap stones.
• ±1 for rank difference of 9 (scaled up/down) for large/extreme handicaps.

Also change Glicko2Entry.copy to add incorporate the error into the rating deviation using root-of-sum-of-squares, which is the usual way to combine standard errors.

[dexonsmith]

Did a sniff test, but not really any analysis of whether the initial defaults are the numbers we actually want.

Let me know if you'd rather default to "off" for these until some analysis is done.

[dexonsmith]

Let me know if you'd rather default to "off" for these until some analysis is done.

I mean, either way there will be analysis to do later, just a matter of the current default on the branch.

[anoek]

What are we looking to improve with this? Comparing the output from with and without this change, the end results are pretty similar, at least on the surface. That leads me to think it's probably fine to merge, but I'm a little hesitant to just click merge without building a bit more intuition about how this is improving results and what, if any, tradeoffs we're making.

[anoek]

From the https://github.com/dexonsmith/goratings/blob/document-ratings-v6/RatingsV6.md the added clarification for the purpose of this is:

Increase effective rating deviation of opponent for larger handicaps and strange komi values.

• Reflects higher uncertainty that opponents' rating is correct, the more the board diverges fom "normal".
• Reduces impact of high handicap games on ratings, enabling a policy decision to allow rated 19x19 games with 10+ stones.
• Reduces impact of reverse-komi games on ratings, enabling a policy decision to allow handicap via reverse komi.

[anoek]

Marking this as draft for the summer so it's not on my "to check on and merge" list until we start revisiting stuff in the fall