fix(terrain): eval-mode env origins for OBJ terrains

[tkevinbest]

Fixes an AttributeError crash at env init when evaluating a policy on an OBJ (load_obj) terrain — e.g. a stairs-climbing policy on a fitted-box mesh.

Terrain spawn origins are computed one of two ways depending on terrain type: procedural terrains store them in Terrain._env_origins, while OBJ terrains never populate that attribute and instead derive per-tile origins from the loaded mesh via _get_load_obj_env_origin_grid(). Separately, TerrainLocomotion._get_env_origins() has two modes: the training branch (randomize_tiles=True) scatters robots across tiles, and the eval branch (randomize_tiles=False) places every robot at tile (0, 0) for deterministic evaluation.

The training branch already fetched origins the right way through sample_env_origins(), which special-cases OBJ terrains. The eval branch predates OBJ-terrain support and reached for self.terrain._env_origins[0, 0] directly — an attribute that doesn't exist for OBJ terrains — so any OBJ-terrain eval crashed on the first _get_env_origins() call.

The fix introduces one accessor and routes both modes through it:

• Adds a Terrain.env_origin_grid property (simulator/shared/terrain.py) that returns the [num_rows, num_cols, 3] per-tile origin grid for either terrain type — _get_load_obj_env_origin_grid() for OBJ, _env_origins for procedural.
• Points both sample_env_origins() and the eval branch in locomotion.py at the property, so eval handles OBJ terrains exactly the way training already does.

Notes for reviewers:

• No behavior change for procedural terrains. The property returns the same _env_origins array the old code used, so existing (non-OBJ) training and eval runs are unaffected.
• Shapes line up at both call sites. Both terrain types yield a [num_rows, num_cols, 3] grid, so the eval branch's [0, 0] (→ one 3-vector) and sample_env_origins()'s fancy index (→ [num_robots, 3]) both work. The procedural grid is float64 and the OBJ grid float32, but both call sites wrap the result in torch.from_numpy(...).to(torch.float), so the dtype difference is normalized away.
• No other call site reads the grid directly. The only external readers were the eval branch (now fixed) and sample_env_origins() (now routed); other _env_origins references are unrelated per-env tensors on different objects.

Verified end-to-end: an OBJ-terrain eval now boots, loads the fitted-box mesh, and records a rollout instead of crashing at init.