[Bug]: SelfCollisionDistance's backward arise a broadcast error

[LeZheng-x]

Prerequisites

• I have searched existing issues and discussions and could not find a duplicate.
• I have read the relevant section of the documentation.
• I can reproduce this on the latest main (or the most recent release).

Bug summary

When joint_positions' horizon >1, using RobotCollisionChecker.get_scene_self_collision_distance_from_joints() to get self_dist, the backward() will arise a broadcast error at curobo/_src/curobolib/cuda_ops/geometry.py, line 101.
The grad_out_distance missing final dimension leads to the error.
g_vec = g_vec * grad_out_distance.unsqueeze(-1) will fix this error.
test_collision.py

Steps to reproduce

from curobo.collision_checking import RobotCollisionChecker, RobotCollisionCheckerCfg
from curobo.scene import Scene
import torch

# Create collision checker for custom pipeline
config = RobotCollisionCheckerCfg.load_from_config(
    robot_config="franka.yml",
    scene_model="collision_test.yml",
    collision_activation_distance=0.01,
)
checker = RobotCollisionChecker(config)

# Query collision at joint configurations
joint_positions = torch.rand((10, 50, 7)).to('cuda:0')  # 10 configurations, horizon 1
scene_distance, self_distance = checker.get_scene_self_collision_distance_from_joints(
    joint_positions
)


# Validate configurations
is_valid = checker.validate(joint_positions)

# Use in custom optimization (differentiable)
joint_positions.requires_grad = True
scene_dist, self_dist = checker.get_scene_self_collision_distance_from_joints(
    joint_positions
)
cost = scene_dist.sum() + self_dist.sum()
print(scene_dist.shape,self_dist.shape)
cost.backward()  # Gradients flow through

Expected behavior

No error happen.

Actual behavior / error output

.venv/lib/python3.11/site-packages/curobo/_src/curobolib/cuda_ops/geometry.py", line 101, in backward
    g_vec = g_vec * grad_out_distance
            ~~~~~~^~~~~~~~~~~~~~~~~~~
RuntimeError: The size of tensor a (65) must match the size of tensor b (50) at non-singleton dimension 2

cuRobo version + commit SHA

0.8.0.post1.dev42+dirty 8e734f3

Installation method

Source — CUDA 13 + PyTorch (uv pip install .[cu13-torch])

Kernel backend

cuda_core (default, runtime compilation)

Python version

3.11.15

PyTorch version (if installed)

2.11.0+cu130 13.0

GPU / driver / CUDA toolkit

RTX 2080 Ti

Operating system

Ubuntu 20.04

Isaac Sim version (if applicable)

No response

Additional context

No response