Support AD on ADTs that have tangent types equal to the primal types

[chenzizhao]

Currently, grad does not work on data types such as List. For example:

ls = (AsList _ [1.,2.,3.])
def sumList ((AsList _ arr) : List Float): Float = sum arr
sumList ls
> 6.
(grad sumList) ls
dex: Not implemented
CallStack (from HasCallStack):
  error, called at src/lib/Autodiff.hs:507:18 in dex-0.1.0.0-RKG8rO926ZDrlxlc3NaeB:Autodiff
  notImplemented, called at src/lib/Autodiff.hs:386:19 in dex-0.1.0.0-RKG8rO926ZDrlxlc3NaeB:Autodiff

Discussion on Friday @apaszke @danieldjohnson @duvenaud For some data types, taking grad with respect to them do not require user-supplied tangent types, since they are the same as the original type. This would be easier to implement than to support grad on any, general data structure.