Add autodiff library node usable in DacePrograms

[orausch]

This adds support for the torch.autograd.backward function, as well as access to .grad of an array.

Calls to torch.autograd.backward inserts a new BackwardPass library node.

Calls to .grad allocate gradient buffers and write to them via these library nodes.

Example:

dace_module = DaceModule(dace_module, auto_optimize=False)

@dace
def train_step(x):
    output = dace_module(x)
    loss = np.add.reduce(output, axis=[0, 1])
    loss.backward()
    return loss

Outputs: Which expands to:

[codecov[bot]]

Codecov Report

Merging #114 (34d6184) into master (febd2dc) will increase coverage by 0.35%. The diff coverage is 96.24%.

@@            Coverage Diff             @@
##           master     #114      +/-   ##
==========================================
+ Coverage   68.32%   68.67%   +0.35%     
==========================================
  Files          60       65       +5     
  Lines        6894     7232     +338     
==========================================
+ Hits         4710     4966     +256     
- Misses       2184     2266      +82     

Impacted Files	Coverage Δ
...l/onnx/op_implementations/cudnn_implementations.py	88.94% <66.67%> (+0.27%)	:arrow_up:
daceml/util/utils.py	71.85% <83.33%> (+0.42%)	:arrow_up:
daceml/autodiff/backward_pass_generator.py	92.55% <91.67%> (+0.13%)	:arrow_up:
daceml/autodiff/analysis.py	92.86% <92.86%> (ø)
daceml/autodiff/library/python_frontend.py	94.94% <94.94%> (ø)
daceml/autodiff/library/library.py	98.58% <98.58%> (ø)
daceml/autodiff/__init__.py	100.00% <100.00%> (ø)
daceml/autodiff/implementations/dace_nodes.py	97.56% <100.00%> (+0.06%)	:arrow_up:
daceml/autodiff/library/__init__.py	100.00% <100.00%> (ø)
daceml/autodiff/library/torch_integration.py	100.00% <100.00%> (ø)
... and 11 more

:mega: Codecov can now indicate which changes are the most critical in Pull Requests. Learn more