Stabilize the single-shift logit transform

[pedrocastanha]

Summary

• rewrite the single-shift logit transform with a numerically stable, algebraically equivalent expression
• preserve the input dtype when constructing the log(2) constant
• add regression coverage for saturated float32 logits and the end-to-end gradient path

Why

The previous expression can overflow for large negative logits and suffer cancellation for positive logits, producing infinities at ordinary saturated float32 values. The new implementation uses logaddexp and remains finite across the tested range from -100 to 100.

Validation

• pytest tests/test_gradients.py -q -k 'single_shifted_theta or single_gradient_is_finite'
• existing param_shift_single simulator gradient tests
• isort, Black, Flake8, and Pyright through pre-commit

Closes #16

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.