Support gradient over function with multiple inputs

[LilithHafner]

This is motivated by https://github.com/SciML/juliatorch/issues/12. It is proof of concept implementation to see if folks here think this would be reasonable to include in ForwardDiff or if it would be better to use a separate package or simply put this in the juliatorch package.

The idea is to automatically pack and unpack inputs into a single array which works with the existing ForwardDiff.gradient API.