clamp xgboost weights to 31

[shuttie]

A fix for issue like this:

08:11:56.589 INFO  a.m.m.r.LambdaMARTRanker$LambdaMARTPredictor - Train/Test split finished: 800295/200074 click-through events
2024-01-17T08:12:01.911Z [WARNING] Your app's responsiveness to a new asynchronous event (such as a new connection, an upstream response, or a timer) was in excess of 100 milliseconds. Your CPU is probably starving. Consider increasing the granularity of your delays or adding more cedes. This may also be a sign that you are unintentionally running blocking I/O operations (such as File or InetAddress) without the blocking combinator.
ml.dmlc.xgboost4j.java.XGBoostError: [08:13:12] /workspace/src/common/ranking_utils.h:355: 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.
Stack trace:
  [bt] (0) /tmp/libxgboost4j12195132296382762716.so(dmlc::LogMessageFatal::~LogMessageFatal()+0x6e) [0x7f626e57ec4e]
  [bt] (1) /tmp/libxgboost4j12195132296382762716.so(xgboost::ltr::NDCGCache::InitOnCPU(xgboost::Context const*, xgboost::MetaInfo const&)+0xcf6) [0x7f626e68f996]
  [bt] (2) /tmp/libxgboost4j12195132296382762716.so(std::shared_ptr<xgboost::ltr::NDCGCache> xgboost::DMatrixCache<xgboost::ltr::NDCGCache>::CacheItem<xgboost::Context const*, xgboost::MetaInfo, xgboost::ltr::LambdaRankParam>(std::shared_ptr<xgboost::DMatrix>, xgboost::Context const* const&, xgboost::MetaInfo const&, xgboost::ltr::LambdaRankParam const&)+0x46b) [0x7f626e96f0bb]

In a case when xgboost is used, and there is a weight > 31, then we emit a warning and clamp it to 31.