Add clamping to [-1, 1] for avoid nans in matrix_to_euler conversion

[tylerlum]

I encountered strange nan issues when converting from quat_wxyz to euler_ZYX like below:

        pos = pose_w[:, :3]
        quat_wxyz = pose_w[:, 3:]
        matrix = quaternion_to_matrix(quat_wxyz)
        euler_ZYX = matrix_to_euler_angles(matrix, "ZYX")
        return torch.cat([pos, euler_ZYX], dim=-1)

I found that the issue was from numerical issues making matrix contain values greater than 1 (e.g., 1 + 1e-6). The domain of torch.acos and torch.asin are [-1, 1], and they return nan if given values outside of this range. Thus, this simply adds clamping to the inputs to these functions to make them robust to these numerical issues.