Reorganize inverse_eltype calculations.

[tpapp]

• remove robust_eltype, replace with _ensure_float
• remove Union{} as a corner case for empty (named)tuples and Constant, use Bool
• clarify API guarantees (inverse_eltype is always <:Real)
• remove _ensure_float from inverse! definition, clarify invariant (eltype is always the same as inverse_eltype)
• incidental: fix various minor code redundancies
• fixes inverse eltype calculation fails for vectors that do not contain scalars #148 (add tests)

[tpapp]

@devmotion: I wonder if this addresses your use case in #132. We fall back to Bool as the narrowest type instead of Union{}, it simplifies the code greatly. The examples from #132 work:

julia> b = Bool[]
Bool[]

julia> sum(b)
0

julia> mean(b)
NaN

julia> std(b)
NaN

If this is satisfactory, and you have time, please also review.

[devmotion]

I worked around this issue for now but IIRC there are still some annoying edge cases. I assume that falling back to a numeric type such as Bool (which is not as invasive as Float64) would generally make it less likely to run into such as issues as in #132. So I'm in favour of this change.

[Ickaser]

As the person using TransformVariables with Unitful: there are at least a couple tests that will be doing inverse transforms with Unitful in a way that resembles my use case, e.g.

TransformVariables.jl/test/runtests.jl

Lines 86 to 89 in 0d70beb

	t1 = TVScale(2u"m") ∘ TVShift(a) ∘ TVExp()
	test_transformation(t1, y -> y > a*2u"m", jac=false)
	t2 = TVScale(1u"s") ∘ TVShift(a) ∘ TVScale(b-a) ∘ TVLogistic()
	test_transformation(t2, y -> (a*u"s" < y < b*u"s"), jac=false)

If you want to be extra careful, you could add an @inferred to the test_transform function, right here:

TransformVariables.jl/test/utilities.jl

Lines 72 to 76 in 0d70beb

	if test_inverse
	x2 = inverse(t, y)
	@test x ≈ x2 atol = 1e-6
	ι = inverse(t)
	@test x2 == ι(y)

[tpapp]

@Ickaser: thanks for the reminder. I am comitted to maintaining support to Unitful and similar approaches. At the same time, I would prefer to error on eg complex arguments that would silently creep through. I added tests for both, now Unitful will definitely work without broadening to ::Number in too many places.