fix(ipynb): read the notebook's own language and a string source

[Lukas (L4XB)]

What

IpynbConverter assumes two shapes that nbformat does not guarantee.

1. Every code cell is fenced as Python.

md_output.append(f"```python\n{''.join(source_lines)}\n```")

An R notebook comes out as:

```python
summary(cars)

The notebook says what it is — `metadata.language_info.name`, and on older notebooks `metadata.kernelspec.language` — and neither was read. A reader, human or model, is told the cell is Python when it is not.

**2. A `source` that is one string loses the title.**

nbformat types `source` as a `multiline_string`: *either* a list of lines or a single string. Both appear in the wild — `nbformat.write` emits lists, but plenty of notebooks are generated by other tooling, and `nbformat.read` accepts both.

`"".join(source)` happens to be correct for both, which is why this went unnoticed. The title scan is not:

```python
for line in source_lines:
    if line.startswith("# "):

Over a string, that iterates one character at a time, and "#".startswith("# ") is never true. Same notebook, two spellings of the same cell:

# source as a list
{"cell_type": "markdown", "source": ["# My Title\n", "\n", "some text"]}  -> title: "My Title"

# source as a string
{"cell_type": "markdown", "source": "# My Title\n\nsome text"}            -> title: None

How

• _fence_language(notebook_content) — metadata.language_info.name, then metadata.kernelspec.language, then "python" as before. A value that is not a usable fence info string (empty, containing a space, containing a backtick) falls through to the default rather than producing a broken fence.
• _source_lines(cell) — splits a string source into lines with splitlines(keepends=True), passes a list through unchanged, and yields nothing for any other type. Joining is unaffected; the title scan now sees lines in both cases.

Raw cells keep their untagged fence — only code cells carry a language.

Test

packages/markitdown/tests/test_ipynb_converter.py — 13 tests.

	main	this branch
tests/test_ipynb_converter.py	4 failed, 9 passed	13 passed

# on main
FAILED test_code_cells_use_the_notebooks_own_language[metadata0-R]
FAILED test_code_cells_use_the_notebooks_own_language[metadata1-julia]
FAILED test_a_string_source_is_read_as_lines
FAILED test_a_source_that_is_neither_a_string_nor_a_list_is_skipped

The nine that pass on both sides are the guards: a notebook with no language metadata still fences as python, a list source renders byte-identically, metadata.title still wins over a heading, raw cells stay untagged, and the three malformed-metadata cases fall back rather than emitting a broken fence.

Full suite: 905 passed, 14 skipped. black (the pinned 23.7.0 from .pre-commit-config.yaml) reports both files unchanged.