fix: handle dict word_list in TextArenaEnv.ta_to_hf()

[tashapais]

Summary

• TextArenaEnv.ta_to_hf() called random.choice(words) assuming word_list is a sequence, but games like TwentyQuestions-v0 expose a categorized dict instead, causing a KeyError on integer indices.
• Fix flattens dict values into a list before sampling, matching how the official Wordle environment on the Hub works.

Test plan

• Instantiate TextArenaEnv(game="TwentyQuestions-v0") and confirm no KeyError on init
• Confirm existing Wordle-v0 behavior is unchanged (list word_list path unchanged)

Closes #1074

---

Note

Low Risk
Low risk: small defensive change isolated to dataset generation logic, only altering how word_list is normalized before random.choice.

Overview
Fixes TextArenaEnv.ta_to_hf() to handle TextArena games where word_list is a dict (e.g., categorized words) by flattening dict values into a list before sampling.

This prevents sampling/index errors while keeping existing behavior unchanged when word_list is already a sequence.

^{Reviewed by Cursor Bugbot for commit a5e7075. Bugbot is set up for automated code reviews on this repo. Configure here.}

[tashapais]

Fixes #1074. The one-line change flattens dict values before sampling, so any TextArena game that uses categorized word lists (TwentyQuestions-v0, etc.) will initialize correctly. @Kannav02 you mentioned wanting to open a PR for this -- happy to hand it off if you'd prefer.

[tashapais]

@willccbb @mikasenghaas ready for review when you get a chance.

[Kannav02]

Fixes #1074. The one-line change flattens dict values before sampling, so any TextArena game that uses categorized word lists (TwentyQuestions-v0, etc.) will initialize correctly. @Kannav02 you mentioned wanting to open a PR for this -- happy to hand it off if you'd prefer.

Its fine, as long as the env gets fixed and works, thanks!