Added stricter state checking for inform7 backend

[christopherzc]

Adds a stronger check when parsing the "score" and "moves" fields for info in the inform7 backend.

https://github.com/pearls-lab/TextWorld/blob/c812d677dafc0728e49ab95cb52a35c0a1876c0c/textworld/envs/wrappers/tw_inform7.py#L138

This is largely needed for using Textworld to train Large Language Models, as it appears that some outputs from the models can cause the inform7 backend to mistake malformed outputs as part of its underlying state tracking, thus leading to an error here.

The try catch here isn't a particularly elegant solution but is robust enough that I haven't had the need to modify it, nor encountered this error again, across a large number of RL training runs.

[MarcCote]

Thanks for the PR. I'm working on fixing the CI stuff, then I'll merge it. If it's urgent, I can merge it sooner.

[christopherzc]

Not really urgent but I'm also trying to see if I can find what specifically is breaking the inform7 backend so a more concrete fix can be implemented. Annoyingly, it doesn't seem to be something that is tied to a single 'bad action' so much as the LLM just inputting junk until something finally breaks in the backend.