feat: implement _update_terrain_curriculum and fix _get_env_origins for heightfield/trimesh

[dcol91863]

Summary

Fixes #81

The _update_terrain_curriculum(env_ids) method was referenced in the docstring of reset_idx but was never implemented. With cfg.terrain.curriculum = True and a heightfield/trimesh terrain, the curriculum was silently a no-op — all robots stayed on the same terrain level forever.

This PR implements the full terrain curriculum loop in a single file (legged_gym/envs/base/legged_robot.py).

---

Changes

_get_env_origins() — heightfield/trimesh branch

Previously the method always created a flat grid and left custom_origins = False, so robots were never actually placed on the heightfield platforms. Now:

• Sets custom_origins = True for heightfield and trimesh mesh types.
• Converts terrain.env_origins (numpy array, shape [num_rows, num_cols, 3]) to a GPU tensor self.terrain_origins for fast indexed lookups.
• Assigns each environment a random terrain type (column index).
• Initialises each environment to a random level (row index) in [0, max_init_terrain_level].
• Populates self.env_origins from the lookup table immediately.

_update_terrain_curriculum(env_ids) — new method

Called at the start of each reset_idx when cfg.terrain.curriculum is True:

1. Computes how far each resetting robot moved from its terrain-platform centre (XY distance).
2. Promotes robots that covered > half a platform length (env_length / 2) to the next harder level.
3. Demotes robots that fell short by one level.
4. Clamps all levels to [0, num_rows - 1].
5. Refreshes self.env_origins[env_ids] from the terrain_origins lookup table so that _reset_root_states places the robot on the correct platform.

The method is a no-op until self.init_done is set, preserving the initial level assignment from _get_env_origins.

reset_idx() — wired up

• Calls _update_terrain_curriculum(env_ids) before robot repositioning when cfg.terrain.curriculum is True.
• Logs extras["episode"]["terrain_level"] (mean across all envs) for training curve monitoring.

Test plan

• Set cfg.terrain.mesh_type = 'heightfield' and cfg.terrain.curriculum = True, train a G1/H1 policy — confirm robots spawn on terrain platforms and terrain_level in tensorboard increases over training.
• Set cfg.terrain.mesh_type = 'plane' — confirm flat-grid path is unchanged.
• Set cfg.terrain.curriculum = False with heightfield — confirm _update_terrain_curriculum is never called and robots still spawn correctly.