refactor: refined sparse reward environment

[neuralsorcerer]

Changes:

• Rewrote SparseRewardEnvironment.observation_space to expose a 4‑dimensional box for agent and goal coordinates and added deterministic seeding on reset.
• Updated continuous‐action step logic to return float tensors, apply time–limit truncation only when the goal is not reached, and defined a bounded BoxActionSpace.
• Improved the discrete environment and validated the action index before converting to the continuous step, keeping predefined action vectors.
• Added tests for ContinuousSparseRewardEnvironment.

Why these changes?

• The observation space previously described only agent coordinates now returning the full agent‑goal state clarifies environment outputs and matches tests.
• The continuous environment now exposes a usable action space and handles termination/truncation correctly, improving compatibility with RL algorithms.

[facebook-github-bot]

@Yonathae has imported this pull request. If you are a Meta employee, you can view this in D78899735.