Fix ustrip broadcasting when range step has different unit than eltype

[sostock]

Fixes #712.

This PR also changes the return type when broadcasting over a StepRange and the conversion factor is a floating-point number: in this case, a StepRangeLen is returned (instead of a StepRange) because it is the better choice for floating-point numbers.

[sostock]

This is bad for performance:

julia> using BenchmarkTools

julia> using Unitful: mJ, eV

julia> r = 1eV:1eV:5eV
(1:5) eV

julia> @btime mJ.($r);
  25.844 ns (0 allocations: 0 bytes) # v1.19.0 (returns a StepRange)
  55.767 ns (0 allocations: 0 bytes) # this PR (returns a StepRangeLen)

We could add methods that call the StepRangeLen constructor directly instead of going through start:step:stop:

julia> simple_mJ(r::StepRange) = StepRangeLen{typeof(mJ(r.start))}(Base.TwicePrecision(mJ(r.start)), Base.TwicePrecision(mJ(r.step)), length(r));

julia> @btime _mJ($r)
  14.013 ns (0 allocations: 0 bytes)

julia> mJ.(r) === simple_mJ(r)
true

This is a simple implementation (maybe the accuracy can be improved in certain cases by setting the appropriate number of trailing zero bits of step.hi and setting the offset so that the reference value is the one with the smallest magnitude), but, at least in the following case, it is actually more accurate than going through start:step:stop:

julia> r2 = 1eV:1eV:500_000eV
(1:500000) eV

julia> last(mJ.(r2))
8.01086714823366e-11 mJ

julia> last(simple_mJ(r2))
8.01088317e-11 mJ

[sostock]

I have now rewritten the function. It is slightly slower than the old version that returned a StepRange, but should be more accurate than the simple version in the previous comment:

julia> @btime mJ.($r)
  25.844 ns (0 allocations: 0 bytes) # v1.19.0 (returns a StepRange)
  33.379 ns (0 allocations: 0 bytes) # this PR (returns a StepRangeLen)