float ranges: for ambiguous length, use round((stop-start)/step)+1

Fixes #20532 by defining which length to use when there's more than
one length that is compatible with a given start:step:stop combo.
This also tightens the tests to check for that specific result.

base/twiceprecision.jl

@@ -147,9 +147,16 @@ function colon{T<:Union{Float16,Float32,Float64}}(start::T, step::T, stop::T)
         end
     end
     # Fallback, taking start and step literally
-    len = max(0, floor(Int, (stop-start)/step) + 1)
-    stop′ = start + len*step
-    len += (start < stop′ <= stop) + (start > stop′ >= stop)
+    lf = (stop-start)/step
+    if lf < 0
+        len = 0
+    elseif lf == 0
+        len = 1
+    else
+        len = round(Int, lf) + 1
+        stop′ = start + (len-1)*step
+        len -= (start < stop < stop′) + (start > stop > stop′)
+    end
     StepRangeLen(TwicePrecision(start, zero(T)), twiceprecision(step, nbitslen(T, len, 1)), len)
 end
 

test/ranges.jl

@@ -346,8 +346,9 @@ for T = (Float32, Float64,), i = 1:2^15, n = 1:5
     lo = hi = n
     while start + (lo-1)*step == stop; lo -= 1; end
     while start + (hi-1)*step == stop; hi += 1; end
+    m = clamp(round(Int, (stop-start)/step) + 1, lo+1, hi-1)
     r = start:step:stop
-    @test lo < length(r) < hi
+    @test m == length(r)
     # FIXME: these fail some small portion of the time
     @test_skip start == first(r)
     @test_skip stop  == last(r)