Implement manual Cholesky for multivariate normal sampling with speedup

[JaLuka98]

This pull request refactors and optimizes the Neyman belt construction and related statistical routines for Gaussian models. The main changes introduce a new 1D Gaussian grid cache abstraction that centralizes and streamlines the generation and evaluation of whitened scan curves, improving both performance and code clarity. The simulation and profile curve logic are updated to use this cache, and sampling is now performed in whitened coordinates for better numerical stability.

Gaussian scan and simulation improvements:

• Added GaussianGridCache1D and related functions (prepare_gaussian_grid_cache_1d, fill_q_samples_for_true_index) in src/npkit/_scan.py to cache whitened means, Cholesky factors, and efficiently generate whitened scan curves and toy samples for 1D Gaussian models.
• Updated the simulation method in GaussianLikelihood to use the cached Cholesky factor for generating samples, ensuring all samples are produced in whitened coordinates. [1] [2]

Neyman belt construction and statistical routines:

• Refactored build_belt and build_belts_from_grid in src/npkit/neyman.py to use the new Gaussian grid cache and whitened scan logic, removing redundant code and improving numerical stability. [1] [2]
• Updated profile_curve_from_likelihood in src/npkit/stats.py to use the Gaussian grid cache for evaluating the profile likelihood curve, and to compute the minimum NLL correctly in whitened coordinates.

Additional integration and API improvements:

• Added imports and minor refactoring to integrate the new scan cache throughout the codebase, including in invert_belt and profile_curve_from_grid. [1] [2] [3]

These changes collectively improve the efficiency, maintainability, and numerical robustness of Neyman belt construction and statistical inference for Gaussian models.