Differentiable parameter learning fix

[taddyb]

What was done:

• Got parameter learning to work for multiple timesteps
• removed some unused data from Github
  • We want only test cases in the data/ dir
• Removed minibatching
• Tested a runoff fix to ensure gradient capture
  • runoff = torch.clamp(ponded_depth - infiltration, min=0)

Testing:

• Tested on small case (744 timesteps)
• Tested on large case (3000 timesteps)
• Tested on many epochs

Notes/Debugging writeups:

Bug:

RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after 
they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or 
autograd.grad(). Specify retain_graph=True if you need to backward through the graph a second time or 
if you need to access saved tensors after calling backward.

Bug in real person speak:

• You have parameters that had a gradient passed through them that were not reset
• Make sure that you free all gradients, and reset intermediate params, before backpropagating a second time

The fix:

• The self.soil_df and self.c attribute variables inside of the dpLGAR().__init__() function were not being updated after backpropagation.
  • These variables rely on updated Van Genuchten parameters to be updated!
• moved this initialization into the set_internal_states() function
• Made sure to reset global parameters too, and the self.ponded_max_depth clone