fix float(::Dual) and add float(::Type{<:Dual})

[stevengj]

Fixes #492, fixes #362, closes #419.

The Base.float function in base supports numeric types, not just values:

julia> float(Complex{Int})
ComplexF64 (alias for Complex{Float64})

and by exploiting this we obtain floating-point promotion consistent with Base.

[codecov-commenter]

Codecov Report

Merging #535 (93a85ef) into master (6b393f4) will increase coverage by 0.10%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #535      +/-   ##
==========================================
+ Coverage   84.83%   84.93%   +0.10%     
==========================================
  Files           9        9              
  Lines         831      830       -1     
==========================================
  Hits          705      705              
+ Misses        126      125       -1     

Impacted Files	Coverage Δ
src/dual.jl	73.07% <100.00%> (+0.34%)	⬆️
src/apiutils.jl	100.00% <0.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6b393f4...93a85ef. Read the comment docs.

[stevengj]

The

Test threw exception
  Expression: dual_isapprox(NESTED_FDNUM ^ PRIMAL, exp(PRIMAL * log(NESTED_FDNUM)))
  DomainError with -1.1858610145546527e18:
  sqrt will only return a complex result if called with a complex argument. Try sqrt(Complex(x)).

makes no sense to me. I actually got a similar error at some point locally, and then it mysteriously disappeared when I re-ran the test — makes me think it is a bug in Julia itself.

Have you seen similar errors prior to this PR?

[KristofferC]

I think I recall seeing something similar to that. It clearly isn't related to the PR.

[mcabbott]

Re NESTED_FDNUM ^ PRIMAL, the rule for ^ attempts to avoid a DomainError from log, and was changed recently. (Both 1.0 and 1 CI are DiffRules v1.2.1, the latest.) Is it possible this is somehow to blame? I don't see how though. https://github.com/JuliaDiff/DiffRules.jl/blob/master/src/rules.jl#L88

Reading the stacktrace more carefully, the error is within isapprox, which calls norm, so it's unrelated to the function actually being tested, which must have run by then.

Note also that I think #530 needs to be merged before tagging anything.

[stevengj]

Is this mergeable, then? If not, what needs to be done?

[dlfivefifty]

We also need Base.big. Can you add that here, or shall I do it in a separate PR?