support sampling with and without replacement

[SpatLyu]

This PR refactors the library sampling logic used in the bootstrap procedure to ensure statistical correctness and improve flexibility.

Key changes

1. Add support for sampling with replacement (bootstrap)

   • When replace_sampling == true, library indices are now sampled with replacement using a uniform distribution
   • This aligns with standard bootstrap methodology for estimating sampling distributions and confidence intervals

2. Retain subsampling (without replacement) as an alternative mode

   • When replace_sampling == false, the existing shuffle-based subsampling strategy is preserved
   • This is useful for library size sensitivity analysis and robustness checks

3. Deterministic behavior when boot == 1

   • When only one bootstrap iteration is requested, the library is taken as the first L indices without randomization
   • This ensures reproducibility and avoids unnecessary overhead

Motivation

Previously, the implementation only supported sampling without replacement, which corresponds to subsampling rather than true bootstrap. This can underestimate variability and lead to biased confidence intervals, especially for nonlinear statistics such as pattern causality measures.

By introducing proper bootstrap sampling:

• The estimated distribution of causality metrics better reflects sampling variability
• Confidence intervals derived from bootstrap replicates become statistically valid

Notes

• The current implementation resamples only the library indices, while keeping prediction indices fixed. This follows the existing workflow, though future improvements may consider joint resampling strategies depending on the underlying data assumptions.