Update default tree-learning parameters and training budgets

[RAMitchell]

This implements the default-parameter update proposed in #12131.

Changes:

• Updates core tree-learning defaults:
  • eta / learning_rate: 0.3 -> 0.1
  • min_child_weight: 1 -> 2
  • subsample: 1.0 -> 0.8
  • colsample_bytree: 1.0 -> 0.8
  • leaves max_depth at 6, following the later issue discussion
• Aligns Python-facing default training budgets:
  • xgboost.train(..., num_boost_round=300)
  • xgboost.cv(..., num_boost_round=300)
  • xgboost.dask.train(..., num_boost_round=300)
  • sklearn fallback DEFAULT_N_ESTIMATORS = 300
• Updates Spark/JVM defaults to match the new tree defaults and numRound=300
• Updates parameter docs, R docs, and Spark tests that asserted the previous sklearn fallback default

One risk to watch in CI is test flakiness from tests that train with implicit defaults but do not set a seed. Since the new defaults enable row and column subsampling by default, any unseeded tests relying on deterministic full-sample training may need explicit seeds or explicit subsample=1 / colsample_bytree=1.

Local checks:

• cmake -S . -B build-cpu -DUSE_CUDA=OFF -DGOOGLE_TEST=ON -DUSE_DMLC_GTEST=ON
• cmake --build build-cpu --target testxgboost -j35
• cmake --build build-cpu --target xgboost -j35
• ./build-cpu/testxgboost --gtest_filter=Param.*:XGBoostParameter.*:Learner.ParameterValidation:GBTree.SelectTreeMethod
• Focused Python pytest slice: 23 passed, 1 skipped