Make autodiff insertion more versatile

[KnutAM]

Two problems are clear from #179

1. The @implement_gradient macro doesn't work if the original function gives anything more specific than AbstractTensor
2. The method doesn't give any error if the output type of the user-specified gradient function is incorrect.

• Export the propagate_gradient function to give the user full control over type specification for dispatching with dual numbers. This makes it possible to solve point 1 above, and this is also documented with an example.
• Throw error for incorrect output from the user-supplied gradient function
• Add tests that dimension mismatches are caught

[codecov-commenter]

Codecov Report

Base: 97.85% // Head: 97.86% // Increases project coverage by +0.01% 🎉

Coverage data is based on head (5c90d4d) compared to base (7a67a82).
Patch coverage: 100.00% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #181      +/-   ##
==========================================
+ Coverage   97.85%   97.86%   +0.01%     
==========================================
  Files          16       16              
  Lines        1211     1219       +8     
==========================================
+ Hits         1185     1193       +8     
  Misses         26       26              

Impacted Files	Coverage Δ
src/Tensors.jl	82.85% <ø> (ø)
src/automatic_differentiation.jl	99.07% <100.00%> (+0.03%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[KnutAM]

Ready for review, @fredrikekre or @KristofferC

[KnutAM]

Found AllTensors today, this could also potentially be used here like

macro implement_gradient(f, f_dfdx)
    return :($(esc(f))(x :: Union{AllTensors{<:Any, <:Dual}, Dual}) = _propagate_gradient($(esc(f_dfdx)), x))
end

Still less specific than f(::Tensor{2,3}), but seems like it should solve #179 while maintaining the syntactic convenience.
Does anyone see any drawbacks to this approach compared to the current implementation?

[KnutAM]

Suggested change

	return :($(esc(f))(x :: Union{AbstractTensor{<:Any, <:Any, <:Dual}, Dual}) = propagate_gradient($(esc(f_dfdx)), x))
	return :($(esc(f))(x :: Union{AbstractTensor{<:Any, <:Any, <:Dual}, Dual}, args...) = propagate_gradient($(esc(f_dfdx)), x, args...))

Note from #197 , requires corresponding update to propagate_gradient