pymjcf: assets dict is too big

I decomposed my model into many pieces of convexhall meshes (about 2k), then tried to load it into pymjcf, with API `mjcf.Physics.from_mjcf_model`, it raised a `ValueError: assets dict is too big`. whether load them with .xml or python's for loop will trigger that Error. But load them with `mujoco.Physics.from_xml_string` goes all right. I don't know if this is a bug in pymjcf or it has some way to avoid. I really loves pymjcf to build my envirment.
[IssuesExample.zip](https://github.com/deepmind/dm_control/files/10979961/IssuesExample.zip)

首先使用 pymjcf.Physics.from_mjcf_model 重现 IssuesExample.zip 中的失败，并将其与 mujoco.Physics.from_xml_string 进行比较。跟踪模型中大约 2,000 个 convex-hull 网格资产的组装方式，并确定为何会达到资产字典限制；当所提供的模型能够成功加载且不出现 ValueError，或已确立有文档记录的 workaround 时，即视为完成。

由索引模型根据 Issue 内容生成。