Add a ForwardDiff extension for dimensionless quantities

[Ickaser]

#682 seemed like low-hanging fruit, so this pull request solves that problem by adding a ForwardDiff extension with a method for convert very much like what the MethodError suggests. It is entirely possible that this method is too specific or not specific enough, but this fixes at least that MWE.

I would love to see this extension eventually allow other functionality, e.g. derivatives with respect to dimensional quantities, but this would be a start at least, and facilitates the case where units are only used inside a user-defined function.

[Ickaser]

@sostock or any other maintainers: is this reasonable? If it just needs some docs or tests, would be happy to contribute something; as it stands, this just fixes #682 .

[sostock]

I have some small comments. A test should be added as well.

[Ickaser]

Should I care about avoiding a SciML solver package like OrdinaryDiffEqRosenbrock as a test dependency? I suspect adding that package would add a lot of compile time to testing, but it would also have the benefit of ensuring that this use case continues to function more broadly.

The basic problem here can be tested (and demonstrated to work) with a relatively simple test, it turns out:
ForwardDiff.Dual(1.0)*u"cm/m" + ForwardDiff.Dual(1.0) is the place where conversion breaks.
I've added that as a test now, so there is something at least now in the test suite making sure this works.