feat: add Randomized LatinHypercube sampler

[Samuel-Vangu]

Description

Summary of changes

• Adds RandomizedLatinHypercube, a randomized Latin Hypercube (LHS) sampler, shaped like Sobol/Halton/RNG — plugs into MonteCarlo.integrate via the existing rng= slot, no changes to MonteCarlo itself.
• torch backend: pure PyTorch implementation (independent random permutation per dimension via argsort on random keys, plus independent uniform jitter within each stratum), generated directly on the target device — no NumPy/SciPy round-trip, same rationale as Halton's torch path (no native LHS engine exists in torch.quasirandom either).
• numpy / jax / tensorflow backends: scipy.stats.qmc.LatinHypercube (already a hard dependency), converted to the requested backend.
• No scramble parameter (unlike Sobol/Halton): this sampler is always the randomized (jittered) LHS variant; there is no deterministic/unrandomized mode to expose.
• size entries (N, dim) accept any strictly positive whole number, not necessarily int (e.g. 4.0 is accepted and converted); non-whole or non-positive values raise ValueError, non-numeric/bool values raise TypeError.
• Adds randomized_lhs_test.py, mirroring sobol_test.py's structure, plus LHS-specific tests (see Test plan).

Resolved Issues

• Addresses Low Discrepancy Sequences for MonteCarlo and VEGAS #140 (MonteCarlo path for randomized LHS; VEGAS + JIT support out of scope here, same as Sobol/Halton)

How Has This Been Tested?

• randomized_lhs_test.py, all backends: analytic accuracy against the test-function collection, seed reproducibility, torch gradient flow
• test_lhs_exact_stratification: exact-equality check (no tolerance) that every dimension is a perfect bijection onto its n strata
• test_lhs_crushes_mc_on_additive_integrand: for a purely additive integrand, LHS's mean squared error is >100x below plain Monte Carlo's, averaged over 30 seeds — Stein's (1987) theoretical guarantee, measured at ~1200x during development
• test_lhs_fixed_seed_is_reproducible, test_lhs_points_in_unit_cube, and 4 input-validation tests (accepts whole-valued floats, rejects non-whole/non-positive/non-numeric/bool)
• ruff format --check .

Related Pull Requests

• feat: add Sobol quasi-Monte Carlo sampler #244 (Sobol) — same issue, same sampler shape/interface, same testing pattern

[Samuel-Vangu]

Hi @gomezzz, this one's ready for CI whenever you get a chance — thanks!

[Samuel-Vangu]

Hi @gomezzz, pushed the ruff format fix — could you approve/trigger the CI run when you
get a chance? Thanks!