Guard against cyclic trees in process_tag/process_element (#256)

[chiliec]

Fixes #256.

Problem

The mutual recursion introduced in 1.2.2 between process_tag and process_element has no cycle guard. A well-formed BeautifulSoup tree is acyclic, but some HTML producers — notably PDF-to-HTML pipelines (the reporter hit this via marker-pdf) — can yield a graph where a descendant node references an ancestor. Converting such input recurses without bound and crashes:

RecursionError: maximum recursion depth exceeded

Because it aborts with an unhandled exception, a single malformed document takes down the whole conversion.

Fix

Thread an optional _visited set of tag id()s through the descent. When a tag already on the current path is re-encountered, return `` to break the cycle instead of recursing into it.

• _visited defaults to None, so existing callers and any custom converters that call process_tag / process_element directly are unaffected.
• A fresh copy is passed per level (_visited | {node_id}), so it only detects a node repeating on the current descent path — a node legitimately reached via two separate sibling branches is still processed each time.

Test

Adds test_cyclic_tree_does_not_recurse, which builds a cyclic tree (a <p> whose contents include its own ancestor <div>), lowers the recursion limit, and asserts conversion completes without RecursionError.

Verification (local, Python 3.11)

• RED→GREEN: with the fix reverted the new test fails with RecursionError; with the fix it passes.
• pytest tests/ — full suite passes (84 tests, up from 83), no regressions.
• flake8 --ignore=E501,W503 markdownify tests (as tox runs it) is clean.