Fix gradients for parameters getting dropped

[DhairyaLGandhi]

Needs SciML/SciMLBase.jl#979 and potentially a change in DifEqBase to pass sensealg in the kwargs in https://github.com/SciML/DiffEqBase.jl/blob/0874ff4ef8df2c74c843ce534e64ab3cdd40efb4/src/solve.jl#L1126

Currently UDEs fail to converge with reverse mode. This is due to dropped gradients for the parameters while solving the adjoint problem. We also definitely need to reduce the number of places where the assumptions SciML adjoints makes differs from the underlying AD. An example is returning the Vector{Vector{Float64}} as the gradient for a sol object instead of the Tangent or NamedTuple (like in Zygote) for the struct we are calculating gradients for.

Checklist

• Appropriate tests were added
• Any code changes were done in a way that does not break public API
• All documentation related to code changes were updated
• The new code follows the
  contributor guidelines, in particular the SciML Style Guide and
  COLPRAC.
• Any new documentation only uses public API

Additional context

Add any other context about the problem here.

[ChrisRackauckas]

Conflicts

[DhairyaLGandhi]

Can largely be closed since the inittype and the parameter gradient accumulation was covered in #1168 (and improved upon).

Only change of note here is the removal of the literal_getproperty, and the way back is called, which can be factored in separately