[python-package] Remove output_margin from XGBClassifier.predict_proba argument list.

[yanboliang]

XGBClassifier.predict_proba is to predict class probabilities, so it doesn't make sense to support output_margin. If users want to output margin, they can just use predict(data, output_margin=True).

Some users report misuse of this argument when calling XGBClassifier.predict_proba at #3308. If users set output_margin=True when calling predict_proba by mistake, it will produce confused and meaningless result.

For example, this is the correct result:
(This is binary classification, each column is the probability of the sample being of a given class)

>>> model.predict_proba(X_test, output_margin=False)[0:10]
array([[0.9545844 , 0.04541559],
       [0.05245447, 0.9475455 ],
       [0.41897488, 0.5810251 ],
       [0.9831998 , 0.0168002 ],
       [0.4119159 , 0.5880841 ],
       [0.31113452, 0.6888655 ],
       [0.9705527 , 0.02944732],
       [0.93274003, 0.06725994],
       [0.11494881, 0.8850512 ],
       [0.6501156 , 0.34988442]], dtype=float32)

And this is wrong result:

>>> model.predict_proba(X_test, output_margin=True)[0:10]
array([[ 4.0454206 , -3.0454206 ],
       [-1.8939297 ,  2.8939297 ],
       [ 0.67301685,  0.32698315],
       [ 5.069422  , -4.069422  ],
       [ 0.64394915,  0.35605082],
       [ 0.20517927,  0.7948207 ],
       [ 4.4952626 , -3.4952626 ],
       [ 3.6295617 , -2.6295617 ],
       [-1.0411582 ,  2.0411582 ],
       [ 1.6195472 , -0.61954725]], dtype=float32)

[codecov-io]

Codecov Report

Merging #3343 into master will not change coverage.
The diff coverage is 100%.

@@            Coverage Diff            @@
##             master    #3343   +/-   ##
=========================================
  Coverage     45.69%   45.69%           
  Complexity      228      228           
=========================================
  Files           166      166           
  Lines         12972    12972           
  Branches        466      466           
=========================================
  Hits           5927     5927           
  Misses         6853     6853           
  Partials        192      192

Impacted Files	Coverage Δ	Complexity Δ
python-package/xgboost/sklearn.py	87.36% <100%> (ø)	0 <0> (ø)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 480e3fd...a776be5. Read the comment docs.

[hcho3]

I went ahead and merged the PR. Thanks!