range uses TwicePrecision when possible (part 2)

[jipolanco]

Follow-up of #44313. See also discussion in #43059.

Fixes the return type of:

• range(start::IEEEFloat; step::Real, length)
• range(; stop::IEEEFloat, step::Real, length)
• range(; stop::Real, step::IEEEFloat, length)

making sure that TwicePrecision is always used in these cases.

[vtjnash]

Can you add some tests for mixed-IEEEFloat calls, since I think there may be some ambiguity errors there:

julia> range(0.0f0, step=1.0e0, length=10)
0.0:1.0:9.0

julia> typeof(ans)
StepRangeLen{Float64, Base.TwicePrecision{Float64}, Base.TwicePrecision{Float64}, Int64}

julia> range(0.0e0, step=1.0f0, length=10)
0.0f0:1.0f0:9.0f0

julia> typeof(ans)
StepRangeLen{Float32, Float64, Float64, Int64}

[jipolanco]

Now ambiguities should be resolved, unless I missed something. I added a bunch of different tests to check this.

TwicePrecision is also now limited to real-valued ranges, so that e.g. complex ranges work correctly again.

For now all of this is done by defining several variants of range_start_step_length and range_step_stop_length. I'm sure there's a more elegant way of doing this, but this solution seems to work and is quite simple to understand.

[vtjnash]

Great work! Thank you