Migrate CuroboPolicy to cuRobo main API (MotionPlanner / DeviceCfg / GoalToolPose)

## Background PR #306 introduced `CuroboPolicy` targeting cuRobo's `0.7.x` API surface (`MotionGen` / `MotionGenConfig` / `WorldConfig`). Validation by @cagataycali on NVIDIA Thor (sm_110, CUDA 13) — see [PR #306 comment](https://github.com/strands-labs/robots/pull/306#issuecomment-4700020819) — confirms the stub-based unit tests (42/42) pass and the policy architecture is solid, but flags that cuRobo's public Python API has been restructured on `main` and the live code paths in `strands_robots/policies/curobo/policy.py` will not import against a fresh install of cuRobo from `NVlabs/curobo` source today. ## API mapping (per Thor validation) | Current code (PR #306) | New cuRobo API on `main` | |---|---| | `curobo.types.base.TensorDeviceType` | `curobo.types.DeviceCfg` | | `curobo.types.state.JointState` | `curobo.types.JointState` | | `curobo.types.math.Pose` | `curobo.types.Pose` (planning needs `GoalToolPose`) | | `curobo.wrap.reacher.motion_gen.MotionGen` | `curobo.motion_planner.MotionPlanner` | | `curobo.wrap.reacher.motion_gen.MotionGenConfig` | `curobo.motion_planner.MotionPlannerCfg` | | `curobo.geom.types.WorldConfig` | Removed — scene goes via `MotionPlannerCfg.create(scene_model=...)` | Shape changes: - `MotionGen.plan_single(start, goal_pose)` -> `MotionPlanner.plan_pose(GoalToolPose, JointState)` - `GoalToolPose` requires 5D tensors `[B,H,L,G,3]` and explicit `tool_frames` - `MotionGenConfig.load_from_robot_config(...)` -> `MotionPlannerCfg.create(robot=...)` - `TensorDeviceType()` -> `DeviceCfg(device=torch.device('cuda:0'))` ## Scope Update the following internal helpers in `strands_robots/policies/curobo/policy.py` to the restructured API while keeping the public surface (`get_actions`, `reset`, `set_robot_state_keys`, the well-known `target_pose` / `target_joints` / `world_update` kwargs) backwards-compatible: - `_build_motion_gen` — switch to `MotionPlannerCfg.create` + `MotionPlanner` - `_build_start_state` — replace `TensorDeviceType` with `DeviceCfg` - `_build_goal_pose` — emit `GoalToolPose` 5D tensors; resolve `tool_frames` from `planner.kinematics` - `_build_goal_joint_state` — same `DeviceCfg` swap; verify `JointState.from_position` import path - `_apply_world_update` — re-route through `MotionPlannerCfg.create(scene_model=...)` since `WorldConfig` is removed ## Validation requirements This is GPU-bound work. Acceptance: 1. Existing 42 stub-based smoke tests still pass (no regressions in the test seam that lets unit tests inject a stub planner without CUDA). 2. The 3 integration tests in `tests_integ/policies/curobo/test_curobo_live.py` pass against current cuRobo `main` on a real CUDA device. Reuse the Franka-7DOF reach-to-pose sanity-check from the Thor validation report as a fourth sanity test. 3. README + module docstring updated with any new install-version guidance once cuRobo cuts a stable release / publishes wheels. 4. Module-level `[curobo]` extra in `pyproject.toml` re-pinned at the first stable cuRobo PyPI release if/when it ships; otherwise leave the extra empty and keep the source-install recipe in the README (current state on PR #306 head). ## Out of scope - The cuRobo `main` API may continue to shift; pin the integration test to a specific cuRobo commit if necessary, with the commit SHA noted in the test docstring. - Goal-set / batched planning (`plan_goalset`, `plan_batch`) was already marked as a separate follow-up in PR #306's "Out-of-scope follow-ups" section. ## Linked - PR #306 — initial CuroboPolicy implementation (closes #301) - #299 — umbrella for non-VLA Policy providers - [Thor validation comment](https://github.com/strands-labs/robots/pull/306#issuecomment-4700020819) by @cagataycali