Add procedural terrain generation for elevation (#28, #19)

[lycocorax]

Fix NaN actions during training (#28) by replacing static elevation terrain with procedural height-field generation (#19). Requires noise == 1.2.2. Elevation RL now defaults to procedural terrain (USD terrain retained for play config).

Procedural terrain generation

• Updated MushrElevationRLEnvCfg to use IsaacLab's height field terrain generator, configurable via NoiseHfCfg
  • Simplex noise-based height field generation implemented in create_noise_hf
  • Procedural road generation (generate_roads = True) in add_roads_v2
  • Terrain is split into 4x4 subterrains with height scaling parameterized by curriculum difficulty

Reward, termination, pose sampling updates

• Fixed NaN in actions by ensuring valid init poses using TerrainBasedPose2dCommandCfg and terrain-based flat_patch_sampling configs
• Cleaned up some unused or legacy terrain-dependent reward terms
• Fixed an issue of circular motion / reward hacking by introducing a curriculum term for decreasing velocity-based reward weight to a lower bound, as well as adding termination penalties

[geiko246]

Is noise==1.2.2 still required? if so, can you put that in the setup.py for wheeledlab_tasks?

Also potentially include a wandb run/report to show expected results.

[geiko246]

i think this method might exist already (not sure if it's in our version but worth checking to reduce lines):
https://isaac-sim.github.io/IsaacLab/main/source/how-to/curriculums.html