[dev-rinchin]

Fix silent training and NumPy sampling parameters

GradientBoosting.fit() currently raises ZeroDivisionError when verbose=0, and treats NumPy sampling scalars such as colsample=np.float64(0.5) as callbacks, producing an unrelated AttributeError. These failures also affect SketchBoost through the shared implementation.

This change allows silent training and accepts Python/NumPy real numeric sampling rates consistently. For example, verbose=0, colsample=np.float32(0.5), subsample=np.float64(0.5) now completes training.

Changes

• Guard iteration progress logging with verbose > 0. Validation metrics, history, and early stopping still run when logging is disabled; positive logging intervals keep their existing behavior.
• Recognize sampling rates with numbers.Real instead of exact Python type comparisons. Accept rates in (0, 1]; reject invalid ranges, NaN/infinities, booleans, arrays, and unsupported objects with parameter-specific errors before training.
• Normalize validated rates to Python float before constructing BaseSampler, preventing NumPy 2 from casting large axis lengths to float16 and overflowing.
• Preserve callable custom samplers implementing the callback lifecycle, including objects that do not inherit Callback.
• Update the GradientBoosting and SketchBoost parameter documentation and add regression coverage.

Validation

• Before the fixes, four selected regression cases reproduced both original exceptions: NumPy float64 for each sampling parameter, and verbose=0 with/without evaluation sets.
• After the fixes, the full test suite passed: 120 passed on Linux with an NVIDIA A100, Python 3.10.22, NumPy 2.2.6, and CuPy 13.6.0.
• Coverage includes Python/NumPy numeric types, invalid rates/types, custom sampler identity and lifecycle, simultaneous row/column sampling in GradientBoosting and SketchBoost, positive logging intervals, validation history, and early stopping with logging disabled.
• Float16 regressions cover both sampling axes at lengths 65,504, 65,520, and 100,000, plus public training on 100,000 rows compared with a Python-float control. The 100,000-element cases for both axes reproduced OverflowError before normalization and pass afterward.
• Predictions and metric histories are compared with control runs using ordinary Python sampling rates and enabled logging, with appropriate floating-point tolerances.
• Changed Python files parse with Python 3.8 syntax; git diff --check passes. Other Python versions and operating systems were not executed.
• Test output contains existing CuPy scatter_add deprecation warnings.

Command:

CUDA_VISIBLE_DEVICES=1 /tmp/pyboost-issue25-venv/bin/python -m pytest -q tests

Fixes #33
Fixes #27