fix: keep blank lines inside fenced code blocks

[kevin (kevin9327)]

Converting a notebook whose code cell holds

def first():
    return 1


def second():
    return 2

produces the same code with the two blank lines between the definitions reduced to one:

'```python\ndef first():\n    return 1\n\ndef second():\n    return 2\n```'

Two blank lines between top-level definitions is what PEP 8 asks for, so this is the common case, not an odd one. The same happens to any <pre> block routed through the HTML converter, and so to DOCX, PPTX, EPUB and RSS content that carries one.

MarkItDown._convert normalizes a converter's result with re.sub(r"\n{3,}", "\n\n", ...). That is right for the ragged whitespace converters leave between blocks, but a fenced code block is the document's own content: its blank lines are data, and rewriting them means the code in the Markdown is not the code in the source.

The collapsing now steps over fenced code blocks. Only a closed fence is treated as a block, so an unterminated one is still normalized, and everything outside a fence collapses exactly as before.

Reproduction

New test file, on unmodified main (packages/markitdown):

$ python -m pytest tests/test_code_block_blank_lines.py -q
FAILED tests/test_code_block_blank_lines.py::test_notebook_code_cell_keeps_its_blank_lines
FAILED tests/test_code_block_blank_lines.py::test_html_code_block_keeps_its_blank_lines
2 failed, 3 passed in 1.56s

with the failure being the rewritten code:

>       assert CODE in markdown
E       AssertionError: assert 'def first():\n    return 1\n\n\ndef second():\n    return 2'
E         in '```\ndef first():\n    return 1\n\ndef second():\n    return 2\n```'

The three tests that pass before and after are the pins: two paragraphs still separate with exactly one blank line, the blank lines around a code block still collapse, and an unterminated fence is still normalized.

With the fix:

$ python -m pytest tests/test_code_block_blank_lines.py -q
5 passed in 1.29s

Verification

Whole suite, before and after — the five new tests are the whole difference:

$ python -m pytest tests -q -k "not test_output_to_stdout and not test_input_from_stdin_without_hints \
    and not test_convert_url and not file_uris and not case_insensitive and not speech"
607 passed, 4 skipped, 39 deselected in 82.30s   # origin/main
612 passed, 4 skipped, 39 deselected in 82.64s   # this branch

(The deselected tests fail on main on this Windows machine for reasons unrelated to the change: console encoding, file-URI paths, and a missing ffmpeg.)

black reports both touched files unchanged.