Add Fisher Information Matrix support and natural-gradient training

[ujohn33]

This PR adds natural gradient descent support to LightGBMLSS by introducing per-distribution Fisher Information Matrix (FIM) scaling of gradients.

What’s included

1. Extend DistributionClass with a natural_gradient flag and a default compute_fisher_information_matrix() method (fallback returns ones).
2. Apply natural-gradient updates in compute_gradients_and_hessians() when natural_gradient=True (gradient scaled by FIM diagonal).
3. Implement closed-form / efficient FIM diagonal computations for several univariate distributions based on the source [Miyamoto et al.] https://arxiv.org/abs/2304.14885 (e.g., Gaussian, Cauchy, Laplace, LogNormal, Logistic, Poisson, StudentT, Weibull, Gumbel), with an optimized branch for response_fn="exp" and a generic autograd-based derivative fallback for other response functions.
4. Add unit tests aligned with the existing test structure to validate:
   • FIM output shapes and positivity/finite values across distributions
   • presence/behavior of the natural_gradient parameter
   • basic end-to-end training sanity for natural_gradient=True

Notes
Natural gradient is not enabled by default; existing behavior is unchanged unless natural_gradient=True.
CRPS currently falls back to standard gradients (a warning is emitted), as natural gradient scaling is implemented for NLL.

[ujohn33]

@StatMixedML @nucflash please have a look once you find time