WIP: Backend-agnostic scene objects + cross-backend domain randomization

[juan-g-bonilla]

Summary

Turns the single, IsaacSim-only rigid-object feature into a general
backend-agnostic scene-asset system, then layers task-independent manager
behavior on top. You can now declare any number of free (dynamic) and static
scene bodies in one top-level scene: config, spawn them on all four backends
(MuJoCo Classic CPU, MuJoCo Warp GPU, IsaacGym, IsaacSim), observe them in the
robot base frame, reset them deterministically, jitter them, and domain-randomize
their physics — from the same declarative config and one unified actor-state API.

What's new

• Scene objects, any count, every backend. Free + static bodies, tri-format
  assets (URDF / USD / XML) auto-selected per backend, and 1→N scene files (one
  file expands into multiple named bodies with per-object fixed/free override).
• Unified per-env actor-state API. Robot, free body, static body, and
  scene-file sub-body are all addressed by name through one 13-D, world-frame,
  name-major/env-minor state API (get/set_actor_states, get_actor_initial_*),
  plus a duck-typed all_root_states view spanning robot + objects on all backends.
• Object observations in the robot base frame (object_pos_b / quat_b /
  lin_vel_b / ang_vel_b); obs dim derived from object count.
• Cross-backend physics DR. Object mass/material/inertia randomizers are now
  registry-driven and run on every backend (was IsaacSim-only, keyed on a dead
  "object" name). MuJoCo DR now also works on the Classic CPU backend.
• Validated, reproducible DR sampling. A single DistributionSpec
  (bounds-first, validated once) + a keyed counter-based TermSampler: the same
  (term, env, episode) draws the same value regardless of term set/order,
  num_envs, or which env subset resets. Typed per-backend material configs name
  only the channels each backend honors (no silent channel drops).
• IsaacSim loader consolidation. URDF and USD spawning collapse into one path
  (object_spawner.py), removing the rigid-props asymmetry and a large amount of
  dead code.
• Minor: initial velocity for scene objects (restored on reset), pose-jitter
  reset term, resolve_asset_path + @holosoma/ package alias, headless-render
  fix (IsaacGym), SimulatorType-based dispatch, fail-loud warnings for
  reward/curriculum tags that match no bodies, and a positive backend-marker CI
  taxonomy + a GPU MuJoCo-Warp image build.

Testing

One real, fully-managed env (real managers, no shims) asserted against the live
simulator; GPU/Isaac backends run the same harness in a subprocess. Coverage:
scene-spawn matrix, cross-backend DR matrix (ranges exaggerated + disjoint from
defaults so a no-op/wrong-backend write is unambiguous), behavioral physics
outcomes, and the unified all_root_states contract. MuJoCo runs in default CI;
IsaacGym/IsaacSim run on the cluster.

Risk

Config-breaking (RobotConfig.object removed) and three module deletions — see
the migration guide. MuJoCo is the only backend in default CI.

---

Migration guide

1. Scene objects: RobotConfig.object → top-level scene=

RobotConfig.object (and the ObjectConfig type) is removed. Declare scene
objects via the new top-level scene: config (a sibling of robot: / terrain:).

# BEFORE — single object hung off the robot config
ExperimentConfig(
    robot=replace(
        robot.g1_29dof_w_object,
        object=replace(
            robot.g1_29dof_w_object.object,
            object_urdf_path="holosoma/data/motions/g1_29dof/whole_body_tracking/objects_largebox.urdf",
        ),
    ),
    ...
)

# AFTER — scene is its own top-level config
ExperimentConfig(
    robot=robot.g1_29dof_w_object,
    scene=scene.g1_29dof_wbt_object_scene,   # from holosoma.config_values import scene
    ...
)

CLI overrides move off the simulator namespace:

# BEFORE
--simulator.config.scene.env_spacing=0.0
# AFTER
--scene.env_spacing=0.0

The largebox asset moved into the shared scene-object tree:

# BEFORE
holosoma/data/motions/g1_29dof/whole_body_tracking/objects_largebox.urdf
# AFTER
holosoma/data/scene_objects/boxes/large_box.urdf

2. Object physics DR config

Object DR terms moved out of terms.locomotion into terms.objects, and material
DR is now a typed per-backend config (each block lists only channels that backend
honors) instead of a flat range dict.

# BEFORE
func="holosoma.managers.randomization.terms.locomotion:randomize_object_rigid_body_material_startup",
params={
    "static_friction_range": [0.1, 0.6],
    "dynamic_friction_range": [0.1, 0.6],
    "restitution_range": [0.0, 1.0],
}

# AFTER
func="holosoma.managers.randomization.terms.objects:randomize_object_rigid_body_material_startup",
params={
    "material": MaterialRandomizationConfig(
        isaacgym=IsaacGymMaterialDR(friction=[0.1, 0.6], restitution=[0.0, 1.0]),
        isaacsim=IsaacSimMaterialDR(static_friction=[0.1, 0.6], dynamic_friction=[0.1, 0.6], restitution=[0.0, 1.0]),
        mujoco=MujocoMaterialDR(sliding_friction=[0.1, 0.6]),
    ),
}
# imports from holosoma.config_types.randomization

base_com_range and inertia ranges likewise move to the typed BaseComRange /
InertiaScale containers. A range can be a bare [lo, hi] (uniform) or a
{kind, low, high, mean, std} spec for gaussian / log_uniform.

3. Renamed / deleted modules — update external imports

Old import	Replacement
holosoma.simulator.mujoco.backends.warp_randomization	holosoma.simulator.mujoco.backends.randomization
holosoma.simulator.isaacsim.usd_file_loader	gone — use simulator.isaacsim.object_spawner
holosoma.simulator.isaacsim.path_utils	gone — use holosoma.utils.path.resolve_asset_path
holosoma.simulator.isaacsim.registry_utils	gone (dead after consolidation)
holosoma.simulator.isaacsim.usd_physics_utils	gone (dead after consolidation)

4. CI test markers (only if you maintain test config)

Tests now carry a positive backend run-target marker
(isaacgym / isaacsim / mujoco / no_sim / requires_inference); selection
switched from -m "not isaacsim" to positive -m "<backend>" with
--strict-markers. Tests outside tests/simulators/<backend>/ need an explicit
pytestmark.