Implement Dirichlet(α) distribution

[jzeuzs]

This PR introduces support for multivariate distributions to peroxide, starting with the Dirichlet distribution.

A separate MVRNG trait was added specifically for N-dimensional random number generation and PDF evaluation. Sampling was implemented using independent Gamma variables.

[jzeuzs]

Hello! Are there any updates on this?

[Axect]

Thanks @jzeuzs, and sorry for the slow review. The overall design is right: MVDist / MVRNG mirror the existing OPDist / TPDist / RNG patterns, the gamma-normalization sampling is the standard algorithm, and the mean/var/cov formulas check out against the test values.

Requesting changes for a handful of small items. Three are inline; the remaining two are branch-level:

• Formatting: CI now runs cargo fmt --all --check on every PR, and this branch currently fails it (trailing whitespace on a few lines in dist.rs and tests/dist.rs, import ordering). A plain cargo fmt fixes all of it.
• Rebase onto current dev: there are no textual conflicts, but the fmt / clippy CI gates were added after this branch diverged, so CI should run against the current base.

Happy to merge once these are in.

[jzeuzs]

@Axect thank you for the review. I have addressed all the feedback.

This PR also opens the way for implementing the Multivariate Normal, Multinomial (Multivariate Binomial), Categorical (Multivariate Bernoulli), and the Multivariate t-Distribution.

I also noticed the Wishart distribution is on the TODO.md. If it is okay with you, I would love to take that on next. Let me know if this PR is good to merge!

[Axect]

Thanks for the quick turnaround, @jzeuzs! I pulled the branch and everything checks out:

• fmt / clippy / tests: cargo fmt --all --check passes, clippy is clean (no errors), and the doctests plus tests/dist.rs all pass.
• The three inline items: all handled. pub use self::MVDist::*; sorts out the doctest scoping, the no_run example compiles now, and moving to ln_gamma keeps ln_pdf finite even for large alpha.
• Numerical check: I cross-checked mean / var / cov / pdf / ln_pdf against scipy.stats.dirichlet on a few alpha vectors, and they match to floating-point rounding (densities agree to ~1e-11, just the ln_gamma vs gammaln gap). The alpha=200 case that used to overflow now returns a correct finite value, and the Gamma-normalization sampler reproduces the theoretical mean over a large draw. Nicely done. 👍

Merging this now, thanks for the contribution!

And yes, I'd be happy for you to take on Wishart next. Looking forward to it!