dirichlet_stability test fails on FreeBSD when compiled with cg_clif

[bjorn3]

cg_clif uses libm functions for most non-trivial float operations rather than x86 SSE instructions. I suspect there is a (1ULP?) inaccuracy in FreeBSD's libm implementation. In the past the same kind of error happened with MinGW, but that seems to have fixed itself at one point.

---- dirichlet_stability stdout ----
---- dirichlet_stability stderr ----
thread 'main' panicked at rand_distr/tests/value_stability.rs:362:5:
assertion `left == right` failed
  left: [0.0002758045685569205, 2.296135759821706e-20, 3.004118281150937e-9, 0.9997241924273248]
 right: [0.00027580456855692104, 2.296135759821706e-20, 3.004118281150937e-9, 0.9997241924273248]

[vks]

I'm not sure how we can fix this. Should we relax the test to accept 1 ULP of difference? Should we report this to FreeBSD upstream?

[bjorn3]

Some float operations have a lower accuracy requirement (eg 1ULP) rather than the 1/2ULP accuracy requirement that basic math operations have. I don't know which operations have that though, but if dirichlet_stability uses one of these, FreeBSD is allowed to return this result and thus the test should be relaxed.