fix: clear stale final observation mask

[TATP-233]

Summary

• clear state.info["_final_observation"] at the start of each NpEnv.step() so it behaves as a per-step event mask
• keep terminal final_observation capture in _reset_done_envs() unchanged for envs that actually finish on the current step
• add regression tests for the stale-mask cases after init_state() and after a subsequent terminal step

Linked Work

• Issue: Closes Bug: _final_observation can stay stale across non-terminal NpEnv.step calls #126
• Milestone:

Validation

• make check
• uv run pytest -m "not slow"
• Additional task-specific validation listed below

Commands actually run:

UV_CACHE_DIR=.uv-cache uv run --no-sync pytest tests/base/test_np_env.py
UV_CACHE_DIR=.uv-cache uv run --no-sync ruff check src/unilab/base/np_env.py tests/base/test_np_env.py
UV_CACHE_DIR=.uv-cache uv run --no-sync ruff format --check src/unilab/base/np_env.py tests/base/test_np_env.py

Impact

• Backend impact: both
• Platform impact: both
• Training effect expected: yes

Artifacts

• W&B:
• benchmark result:
• video / screenshot:
• ONNX / checkpoint:

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed
• Linked the driving issue
• Noted any follow-up work explicitly

Follow-up work: none.