Add detailed tests + structural properties implementation

[loreloc]

• Added some tests for the region graphs
• Made tests about compilation independent of the region graph module
• Added a couple of tests having ground truths computed by hand

Current coverage is ~73%.

• Added the previously missing structural properties checks: is_compatible, is_structured_decomposable, is_omni_compatible. Note that these properties are stronger than the respective ones that allow for rearrangements of product units as to match particular variables partitionings. A more complicated implementation might be required in the future to deal with such product rearrangements.
• Added __repr__ implementations as to print useful debug information such as shapes, e.g., printing tensorized circuits now yields

TorchCircuit(
  (0): TorchCategoricalLayer(
    folds: 5  channels: 1  variables: 1  output-units: 2
    input-shape: (5, 1, -1, 1)
    output-shape: (5, -1, 2)
    (probs): TorchParameter(
      shape: (5, 2, 1, 2)
      (0): TorchTensorParameter(output-shape: (5, 2, 1, 2))
    )
  )
  (1): TorchCPTLayer(
    folds: 2  arity: 2  input-units: 2  output-units: 2
    input-shape: (2, 2, -1, 2)
    output-shape: (2, -1, 2)
    (weight): TorchParameter(
      shape: (2, 2, 2)
      (0): TorchTensorParameter(output-shape: (2, 2, 2))
    )
  )
   ...
)

Bug fixes:

• Fixed a bug in the Circuit.from_region_graph method that introduced dummy product layers having arity=1.
• Fixed a bug in the circuit optimization routine of the torch compiler that allowed some optimization patterns being wrongly matched when the tensorized circuit structure was a DAG (e.g., QuadGraph).

Minor changes:

• Suggested by @gengala : Categorical and Gaussian layers are now univariate, i.e., they can only accept a scope with one variable (but still multiple channels). This simplifies the integration of a proper set of variables A LOT, while also simplifying the overall symbolic representation by pruning possible different ways to build the same circuit.

Closes #213 by refactoring the region graph data structure. Closes #57 as it is not an informative issue and will never be closed. Closes #228 as we now have basic implementations of the structural properties

[codecov[bot]]

Codecov Report

Attention: Patch coverage is 76.56968% with 153 lines in your changes missing coverage. Please review.

Project coverage is 72.27%. Comparing base (faf0c0c) to head (e0859be). Report is 36 commits behind head on main.

Files with missing lines	Patch %	Lines
cirkit/templates/region_graph/graph.py	77.21%	29 Missing and 7 partials :warning:
cirkit/templates/region_graph/algorithms.py	86.43%	15 Missing and 20 partials :warning:
cirkit/symbolic/circuit.py	64.44%	24 Missing and 8 partials :warning:
cirkit/backend/torch/graph/modules.py	58.18%	22 Missing and 1 partial :warning:
cirkit/symbolic/operators.py	0.00%	4 Missing and 2 partials :warning:
cirkit/backend/torch/compiler.py	54.54%	3 Missing and 2 partials :warning:
cirkit/backend/torch/layers/inner.py	0.00%	2 Missing and 2 partials :warning:
cirkit/symbolic/functional.py	25.00%	1 Missing and 2 partials :warning:
cirkit/backend/torch/graph/folding.py	80.00%	1 Missing and 1 partial :warning:
cirkit/backend/torch/parameters/nodes.py	60.00%	2 Missing :warning:
... and 4 more

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #255 +/- ## ========================================== + Coverage 69.08% 72.27% +3.19% ========================================== Files 53 45 -8 Lines 4392 4426 +34 Branches 594 631 +37 ========================================== + Hits 3034 3199 +165 + Misses 1219 1057 -162 - Partials 139 170 +31 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.