Handling case where all molecules are invalid in property optimization

[jannisborn]

Currently (with the current code on master), if one runs the property optimization and the model produces only one unique molecule, we get this error:

torchdrug/torchdrug/tasks/generation.py", line 147, in reinforce_forward
    raise ValueError("Generation results collapse to singleton molecules")
ValueError: Generation results collapse to singleton molecules

Moreover, when we have only invalid molecules, we get another error:

torchdrug/torchdrug/tasks/generation.py", line 809, in reinforce_forward
    if graph.num_nodes.max() == 1:
RuntimeError: max(): Expected reduction dim to be specified for input.numel() == 0. Specify the reduction dim with the 'dim' argument.

Conceptually, the error is basically the same as reported in #83 (which was fixed in #84). But this time, the affected part is the property optimization, not in the training.

SOLUTION:
Basically, what I'm doing in this PR is imitating the two return variables of reinforce_forward whenever the number of generated molecules is 0 or 1.

Please have a look @KiddoZhu

[jannisborn]

Hi @KiddoZhu, it would be great if you could have a look at the PR