deepcopy fails to copy custom node attributes from document tree in nodes.py:inline_all_toctrees

Describe the bug

I've been working on an extension to add an rst directive that lets me specify a path to a csv file and a caption, then generates an HTML or LaTeX table from that directive.

when the visit_csv_table_node_html method is called (for HTML/EPUB), I am able to access all of the attributes I added to CSVTableNode. The generated HTML looks correct in both output formats.

when the visit_csv_table_node_latex method is called, an AttributeError is raised when I try to access node.data_path.

I traced this down with the debugger to the inline_all_toctrees routine in sphinx/util/nodes.py where the doctree is going through a deepcopy (https://github.com/sphinx-doc/sphinx/blob/master/sphinx/util/nodes.py#L468). Before this copy operation, I am able to access the data_path attribute in the doctree. After this copy, I get an attribute error:

commenting out that line seems to have resolved my issue and I can now generate PDF files (with correct output from what I can tell). Could this be user error on my part, or is there something wrong with the implementation of deepcopy for the nodes.document type?

as a note, the actual table generation code for HTML/LaTeX hasn't been implemented yet - I was working on the plumbing when I found this, hence the very simple elements being added to the "body" object.

index.rst:

index.rst.txt

conf.py:

conf.py.txt

csv_table.py (my extension):

csv_table.py.txt

How to Reproduce

1. run the sphinxdoc/sphinx-latexpdf image
2. using my index.rst, conf.py, and csv_table.py build the project with make latex/make html

Environment Information

Using the official sphinx-latex image from dockerhub: 
docker.io/sphinxdoc/sphinx-latexpdf:latest
Image ID: a0591e845b10

Sphinx extensions

Additional context

No response

Start with the attached index.rst, conf.py, and csv_table.py files, then reproduce the HTML and LaTeX builds using the sphinxdoc/sphinx-latexpdf image. Read sphinx/util/nodes.py at inline_all_toctrees and inspect the deepcopy of the document; done means the custom node's data_path remains available during LaTeX processing without breaking the existing HTML output.