Fix uniform sample weight knot placement

[jxudata]

Summary

Fixes weighted knot placement so uniform nonzero sample_weight values preserve the same interpolated quantiles as the unweighted path.

Root Cause

The weighted quantile path used a step-CDF lookup for all weighted calls. Passing sample_weight=np.ones(n) therefore changed knot placement relative to sample_weight=None, which could change the spline basis and fitted predictions for a semantic no-op.

Changes

• Validate sample-weight shape before fitting.
• Validate _weighted_quantile weight shape directly.
• Ignore zero-weight samples for knot placement.
• Delegate to np.quantile for uniform nonzero weights to preserve NumPy interpolation.
• Add regression coverage for uniform-weight quantiles, knot placement, and 2-D sample-weight validation.

Validation

• Local: PYTHONPATH=src pytest -> 133 passed, 1 xfailed
• GitHub Actions: Python 3.9, 3.10, 3.11, and 3.12 all passed