Fix graphgym activation function dictionary, now not including instances anymore

[fjulian]

Until now, the act_dict used by the graphgym part of this library contained instances of activation functions that were used to build models. This is okay for the activation functions that do not have parameters. However, in the case of PReLU, there is a parameter. This lead to the problem that every PReLU which is built using the graphgym infrastructure shared a single parameter, e.g. when PReLUs are used in multiple layers, or when PReLUs are used in several models that are trained alternatingly. The shared parameter caused unexpected changes to a network that was thought to be frozen while another network was trained.

The solution to this problem proposed in this PR is to include activation function constructors in the act_dict instead of instances, and create a new instance every time an activation function is used in a model.

[fjulian]

Interestingly, now the test test_run_single_graphgym in test/graphgym/test_graphgym.py fails, because the number of parameters of a specified model (23883) does not correspond to the number expected by the test (23880) anymore. I assume this is because the model contains 4 PReLUs, which have now separate parameters with my proposed changes.
So either it was desired that the model's PReLUs share their weights, in which case we would need to come up with a more flexible solution than what I'm proposing here, or it was not desired and my changes fix the bug. In the latter case, the test can be fixed by updating the magic number I mentioned above. In either case, I wonder how this number was obtained in the first place.

[rusty1s]

Thanks @fjulian, let's fix/adjust the test.

[fjulian]

My pleasure. Just fixed them.

[codecov]

Codecov Report

Merging #4978 (022edec) into master (9b129b8) will decrease coverage by 0.01%.
The diff coverage is 77.27%.

@@            Coverage Diff             @@
##           master    #4978      +/-   ##
==========================================
- Coverage   82.74%   82.73%   -0.02%     
==========================================
  Files         330      330              
  Lines       17924    17938      +14     
==========================================
+ Hits        14832    14841       +9     
- Misses       3092     3097       +5     

Impacted Files	Coverage Δ
torch_geometric/graphgym/models/act.py	80.00% <76.19%> (-20.00%)	⬇️
torch_geometric/graphgym/models/layer.py	73.11% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9b129b8...022edec. Read the comment docs.

[rusty1s]

LGTM.