fallback randn/randexp for AbstractFloat

[stevengj]

This PR adds a fallback randn using the Marsaglia polar method (a Box–Muller variant) and a fallback randexp using -log(rand()) so that these two functions work for any AbstractFloat type implementing rand(rng, T).

As noted in #96 (comment), this gives us support for packages like DecFP.jl (closes JuliaMath/DecFP.jl#80) and DoubleFloats.jl for free.

Also closes #17629 for BigFloat (though a specialized method ala #35111 should be faster).

@andreasnoack, is there a way to test this using https://github.com/JuliaRandom/RNGTest.jl or similar?

[stevengj]

(It's possible that the original Box–Muller algorithm is better here than the Marsaglia variant. It depends on the cost of a cos call compared to a 22% probability of 2 additional rand calls?)

I did some benchmarks with Float64, Dec64 (from DecFP.jl), and BigFloat, and the Marsaglia algorithm always seems to be faster (on average) than Box–Muller. The exception to the rule is Dec128, where Marsaglia is 20% slower for some reason.

[stevengj]

The SparseArrays test checks that randn(BigFloat) throws a MethodError: https://github.com/JuliaSparse/SparseArrays.jl/blob/97189f3f36cce0a44210853bccfa4163e30678f3/test/sparse.jl#L3001

How can I update this test given that this is in another repo?

[DilumAluthge]

Hmmm. Maybe first make a PR to SparseArrays.jl to remove that test? And then bump SparseArrays.jl on Julia master.

[stevengj]

Note that this is a minor change, but a MethodError seems to be what is really desired here (and is what we normally throw in methods we've only defined to prevent infinite dispatch loops).

(Otherwise, some downstream tests in SparseArrays fail because randn now falls through to rand and throws a different error, so this seems like the least-breaking change.)

[vtjnash]

need to address CI error:

Error in testset SparseArrays/sparse:
Test Failed at C:\buildbot\worker-tabularasa\tester_win64\build\share\julia\stdlib\v1.9\SparseArrays\test\sparse.jl:2976
  Expression: sprandn(T, 5, 5, 0.5)
    Expected: MethodError
  No exception thrown

[DilumAluthge]

Needs #44731

[stevengj]

Rebased, so it should include the fix #44731 now.

[stevengj]

Okay to merge?

[stevengj]

Unrelated CI failures.