More stringent range testing

[timholy]

Turns a @test_skip into a @test. This may have been passing since #23194 but was never turned into a requirement. This change removes the single largest contribution to "total broken tests" in the whole test suite.

[LilithHafner]

Error in testset ranges:
--
  | Test Failed at C:\buildkite-agent\builds\win2k22-amdci6-4\julialang\julia-master\julia-3d18b7b681\share\julia\test\ranges.jl:890
  | Expression: stop ≈ last(r)
  | Evaluated: -6.67572f-5 ≈ -6.681681f-5

[timholy]

Yeah, there seem to be a few edge cases left. Would could allow a small failure rate, or just close this. Thoughts?

[simonbyrne]

Why is it failing only on some platforms? Is it not deterministic?

[simonbyrne]

An example case I found:

julia> (start, delta, stop)  = -2.8345144f0, 0.9447948f0, -0.00012993813f0
(-2.8345144f0, 0.9447948f0, -0.00012993813f0)

julia> r = start:delta:stop
-2.8345144f0:0.9447948f0:-0.00013005733f0

julia> last(r) ≈ stop
false

The problem here is that

julia> start + 3*delta  # intermediate values in Float32
-0.00012993813f0

julia> Float32(Float64(start) + 3*Float64(delta)) # intermediate values in higher precision
-0.00013005733f0

This particular case uses the fallback branch, since the rational part fails:

julia> num, den = Base.rat(delta)
(599, 634)

julia> Float32(num/den)
0.94479495f0

The question here is what criteria we should use for determining delta and stop?

The fix for now is to adjust the tolerance.

[timholy]

LGTM, thanks @simonbyrne

[vtjnash]

Seen on CI:

The global RNG seed was 0xf5fc6174bf915507070615bf2c22eaf4.
Error in testset ranges:
Test Failed at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-macmini-aarch64-4.0/build/default-macmini-aarch64-4-0/julialang/julia-master/julia-d7b2062924/share/julia/test/ranges.jl:890
  Expression: ≈(stop, last(r), atol = eps((n - 1) * Δ) + eps(stop))
   Evaluated: -1.6152623f0 ≈ -1.6152625f0 (atol=1.7881393e-7)