Return current observations from the RSL-RL wrapper

[sylvesterkaczmarek]

Summary

RSL-RL's VecEnv.get_observations() contract is to return the current observations. RSLRLBraxWrapper.get_observations() currently delegates directly to reset(), so querying observations resets every wrapped environment.

This matters because RSL-RL queries observations when constructing OnPolicyRunner and again at the start of learn(). With the current wrapper, the second query resets the environments again immediately before rollout.

This change:

• extracts the existing observation conversion into _current_observations();
• preserves initialization by calling reset() when env_state is still None;
• returns observations from the existing env_state on subsequent get_observations() calls;
• keeps explicit reset() behavior unchanged.

RSL-RL references:

• VecEnv.get_observations(): https://github.com/leggedrobotics/rsl_rl/blob/main/rsl_rl/env/vec_env.py
• OnPolicyRunner observation queries: https://github.com/leggedrobotics/rsl_rl/blob/main/rsl_rl/runners/on_policy_runner.py

Validation

• Added a regression test verifying that get_observations() does not call reset() once the environment is initialized.
• Added a complementary test verifying that an uninitialized wrapper still initializes through reset().
• Audited the final branch diff: 2 files, +61/-5 lines, with no unrelated changes.
• Squashed to one focused commit: 522453f80c66e8099419fcd1c3e85d4778fd8496.
• Branch is based on current upstream main at e74217bb89c77a74ba02e4789263991864375799.
• Full tests were not run locally in this environment. The focused regression tests are included for upstream CI validation.