Add --handicap-rank-difference-{small,19x19} options

[dexonsmith]

Add options to compute the effective handicap rank difference, incorporating board size, komi, and the point value of handicap stones.

This hits a math domain error when tallying because the head start for black is sometimes enormous, such as 8 stones on a 9x9 board.

Processing OGS data
   2,739,321 /   30,811,698 games processed.  192.3s remainingTraceback (most recent call last):
  File "/Users/dexonsmith/Repos/online-go/goratings/analysis/analyze_glicko2_one_game_at_a_time.py", line 99, in <module>
    tally.add_glicko2_analytics(analytics)
  File "/Users/dexonsmith/Repos/online-go/goratings/analysis/util/TallyGameAnalytics.py", line 102, in add_glicko2_analytics
    self.prediction_cost[size][speed][rank][handicap] += - math.log(result.expected_win_rate if black_won else 1 - result.expected_win_rate)
                                                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: math domain error

Since the options are off-by-default, it'd fine to land as-is and then decide what to do. Or we could fix first.

Here are the choices for what to do (either before or after this lands):

• Update the tallying code to check for 0 before calling math.log().
• Clamp effective rank between (e.g.) 40kyu and 15dan
• Stop rating games with a handicap rank difference bigger than (e.g.) 9
• Some combination above the above.

Relates to #45. Replaces #46.

[anoek]

Stop rating games with a handicap rank difference bigger than (e.g.) 9

I suspect this is the answer, especially for some of those games where the effective rank difference is computed to be something like 52 or whatnot. When that happens your going to have the effective player ratings be something like 1000 vs 5000, and that's fine so long as the 5000 wins (which should net basically 0 points to both parties), but if they do something like resign that game for whatever reason, all of a sudden you're going to extremely tank one and vault the other a ridiculous amount, not to mention probably jacking up their RD's high so as to further increase their rating volatility going forward. I suspect this is what's leading to those math domain errors.

So yeah I don't think considering those types of games is going to do anything but harm the rating pool. If we want to experiment with considering some past 9 but maybe not all the way to 50+, that might be something to consider, but intuitively I wouldn't expect many of those games to be beneficial.