Accept any AbstractVecOrMat when sampling from an MvNormal

[devmotion]

Fixes #2086.

_rand! for MvNormal was defined for VecOrMat === Union{Vector,Matrix}, so any other AbstractMatrix — a view, most commonly — missed it and fell back to the generic eachvariate method in genericrand.jl, which draws column by column. That replaced one randn! and one BLAS-3 unwhiten! with size(x, 2) scalar randn loops and size(x, 2) BLAS-2 unwhiten! calls.

This widens the signature to AbstractVecOrMat and drops the separate scalar-loop method for AbstractVector, whose comment — "randn! only works for Array, but not generally for AbstractArray" — no longer holds: randn! has a generic AbstractArray method on every Julia version this package supports (julia = "1.10"), and MvNormalCanon and GenericMvTDist already accept an AbstractMatrix and rely on it.

For a 10×1000 draw with a dense covariance matrix:

	before	after
MvNormal, Matrix	21.416 μs	21.416 μs
MvNormal, view	115.583 μs	43.375 μs
MvLogNormal, view	149.625 μs	76.417 μs

MvLogNormal follows because it takes an AbstractVecOrMat and forwards into MvNormal. The view path now sits level with MvNormalCanon's (45.875 μs), which already had the wider signature.

Behaviour:

• Vector and Matrix arguments keep the body they already dispatched to; a seeded Matrix draw hashes identically before and after.
• For other AbstractVecOrMat the drawn normals are unchanged, since the generic randn! walks eachindex in the same column-major order the per-column loop did. Only unwhiten! moves from size(x, 2) BLAS-2 calls to one BLAS-3 call, which is bit-identical for a PDiagMat and differs at the rounding level for a PDMat (maxdiff = 8.9e-16 on a 6×40 draw).
• Test.detect_ambiguities(Distributions; recursive = true) stays at 0. (MvNormal, AbstractVecOrMat) is strictly more specific than (Sampleable{<:ArrayLikeVariate}, AbstractArray) in both arguments.
• PDMats' unwhiten! is already declared for AbstractVecOrMat throughout, so no companion change is needed.
• MvLogitNormal is affected beyond the view case. Its _rand! passes @views _drop1(x) into rand!(rng, d.normal, …), i.e. always a SubArray, so even a plain Matrix draw moves from the column loop to the bulk path and its output changes at the rounding level in the common case.

Since the change is behaviour-preserving up to rounding, the testset can only do so much. The one assertion that fails on master is that sampling into a view of a whole matrix gives exactly the same samples as sampling into the matrix, and it fails there through the rounding difference between one BLAS-3 unwhiten! and one BLAS-2 call per column (maxdiff = 4.4e-16 for the covariance matrix used). The remaining assertions guard against breakage rather than proving that the bulk path is taken: a view of some columns is filled like a matrix of the same size and nothing outside it is written, for a PDMat, a PDiagMat and a ScalMat, for a view that is not strided and hence cannot use BLAS in unwhiten!, for a single column in place of the removed AbstractVector method, and for a view without columns.

That exact equality holds only because StableRNG has no bulk randn!. The ones in Random are defined for MersenneTwister/Xoshiro and Array{Float64}, and a view satisfies neither, so with those RNGs a Matrix and a view draw different streams by design, both before and after this PR. This is also the residual performance gap between the two, and it is a Julia-level matter that is not addressed here.

---

This pull request was written with the assistance of generative AI (Claude Code).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.76%. Comparing base (54720dc) to head (0d6ea6b).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2087      +/-   ##
==========================================
+ Coverage   86.75%   86.76%   +0.01%     
==========================================
  Files         149      149              
  Lines        8881     8888       +7     
==========================================
+ Hits         7705     7712       +7     
  Misses       1176     1176              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.