default_rng streams matched NumPy 1.26-2.5, except multivariate_normal

Seed 12345, Linux x86-64 PyPI wheels of NumPy 1.26.4, 2.0.0, 2.1.0, 2.2.6 and 2.5.4: random, normal, standard_normal, exponential, gamma, beta, integers, choice (without replacement, and with p), permutation, shuffle, binomial, poisson, multinomial and hypergeometric gave bit-identical arrays on all five, as did legacy RandomState.normal. multivariate_normal (2x2 covariance) with the default method='svd' was identical on 1.26.4 through 2.4.0 (2.3.5 and 2.4.0 also tested) but changed in 2.5.4. 51% of 200k values differ, by at most 8.9e-16 absolute, consistent with an unchanged stream and a last-bits change in the SVD of the covariance (bundled OpenBLAS 0.3.30 in 2.4.0, 0.3.34 in 2.5.4). method='cholesky' and method='eigh' stayed bit-identical on 2.2.6, 2.3.5, 2.4.0 and 2.5.4. For regression fixtures, compare multivariate draws with a tolerance; method='cholesky' was stable here, but it also goes through LAPACK, and I only tested a 2x2 matrix.