XGBoost 2.0.0 breaks tests

[ksaur]

XGB 2.0.0 was released last week (Sept 12, 2023). This causes 19 tests to fail.

Many errors are related to NDCG ("NDCG is now the default objective function.")

[ksaur]

Before, I believe they were defaulting to objective="rank:pairwise" but that doesn't seem to fix this.
Also, we'll need an upstream fix from onnxmltools for the 1 onnx XBG error.

Started troubleshooting at #733

Pinning this for now.

[ksaur]

onnxmltools was updated but let's wait for their release, then do ours

[ksaur]

With the onnxmltools patch, we now get some new errors:
==== 19 failed, 516 passed, 105 skipped, 115 warnings in 294.32s (0:04:54) ====

Some of them are related to the ndcg change

FAILED ....Check failed: label_is_valid: Relevance degress 
must be lesser than or equal to 31 when the exponential NDCG gain function is used. 
Set `ndcg_exp_gain` to false to use custom DCG gain.

but many are just outright wrong: Mismatched elements: 100 / 100 (100%) ...

[ksaur]

The problem is related to convert_sklearn_xgb_regressor's

hummingbird/hummingbird/ml/operator_converters/xgb.py

Line 133 in 1ae1152

tree_infos = operator.raw_operator.get_booster().get_dump()

[ksaur]

With xgb==1.7.6, for get_dump() we have '0:[f0<0.84472543] yes=1,no=2,missing=1\n\t1:leaf=-0.0600000024\n\t2:leaf=0.100000009\n',

whereas with xgb==2.0.0 we have '0:leaf=2.36637643e-09\n' from get_dump()

Then at:

hummingbird/hummingbird/ml/operator_converters/xgb.py

Line 81 in 1ae1152

return TreeParameters(lefts, rights, features, thresholds, values)

the values for TreeParameters(lefts, rights, features, thresholds, values) are incorrect

[ksaur]

Since it seems they likely changed something internally (now it appears there is a one-node tree rather than a three-node tree), this will take a bit more time. Let's not block the release on this for now.

[ksaur]

TODO: look at onnx/onnxmltools#597 more closely

[ksaur]

Going forward we can look at #764 as a starting point