Fix: resolve NaN gradients in RQS out-of-bounds inverse pass

[cschoenle]

Fix: Resolve NaN gradients for out-of-bounds inputs in RationalQuadraticSpline

Closes #239

Description

This PR resolves a regression introduced in v18.0.0 where evaluating the RationalQuadraticSpline on out-of-bounds inputs caused NaNs in the reverse-mode gradients for the inverse pass, poisoning the gradients for the entire batch.

The Issue:
Although the final output safely defaults to the identity mapping via jnp.where for out-of-bounds inputs, the mathematical operations were still being evaluated on the raw, unclipped variables. In the inverse pass, this caused jnp.sqrt to evaluate a negative discriminant for out-of-bounds data. This resulted in NaN values that propagated through the unselected branches during backpropagation.

The Fix:

• Inverse Pass (The Bug): Restored numerical safety by applying jnp.clip to y to bound it to self.interval strictly before computing the bin index k and intermediate variables. The original, unclipped y is preserved solely for the final jnp.where fallback.
• Forward Pass (Safeguard): Proactively applied the exact same clipping logic to x in transform_and_log_det. While the forward pass was not strictly producing NaNs (as the polynomial math happens to avoid undefined domains), applying this ensures architectural consistency and protects against potential overflows.

Testing:

• Added a new parametrized unit test test_RationalQuadraticSpline_out_of_bounds_grad_nan in test_rational_quadratic_spline.py.
• The test simulates a noisy initialized state and evaluates both the forward and inverse passes on out-of-bounds inputs (50.0), asserting that no NaNs exist in the resulting gradient leaves. Both passes test cleanly.