Preserve ranking-loss gradients with weighted log-sigmoid - #4116

LambdaLoss computes a weighted sigmoid in probability space, then clamps it before taking the logarithm. For strongly misordered documents, this caps the loss at -log(eps) and gives the model zero ranking gradients. RankNetLoss inherits the same behavior. With scores [0, 1000] and relevance [1, 0], the natural-log RankNet loss is about 23 instead of 1000.

Compute the weighted log-probability directly with logsigmoid, and convert to binary logarithms afterward. This preserves the pair selection, ranking weights, and mean reduction while retaining gradients for badly misranked documents.

Validation: CPU tests import the complete native source package and run the production losses with a deterministic trainable score model at the model-input boundary. The 50 new cases compare losses and gradients against independent float64 binary cross-entropy references, covering both losses, both logarithm bases, multiple sigma values, large positive/negative score gaps, and variable-length padded batches. Before the change: 16 failed, 34 passed. After: 50 passed. The final targeted run included the existing ListMLE numerical regression cases: 62 passed, 16 deselected. Ruff lint and formatting checks pass for both changed files. No pretrained-model or CUDA training run was performed.