feat(detect): index Jupyter notebooks with kernel-language fenced code blocks

[Copilot]

Ports notebook indexing from Graphify-Labs#1498 with the core fix: fenced code blocks use the actual kernel language from notebook metadata instead of hardcoded code.

Changes

graphify/detect.py

• NOTEBOOK_EXTENSIONS = {'.ipynb'} — kept separate from CODE_EXTENSIONS/DOC_EXTENSIONS since notebooks go through sidecar conversion
• ipynb_to_markdown(path) — converts cells to markdown; resolves fence language via metadata.language_info.name → kernelspec.language → "code"; handles source as either str or list[str] (both valid per Jupyter spec)
• convert_notebook_file(path, out_dir) — sidecar logic mirroring convert_office_file(); uses content-equality check so output-only re-runs don't churn sidecar mtime
• classify_file() — .ipynb → DOCUMENT
• detect() — converts notebooks to markdown sidecars before indexing

tests/test_detect.py — 14 new tests covering language resolution priority, mixed cells, mtime stability on output-only changes, detect()/detect_incremental() integration, and string vs. list source formats.

Language resolution

meta = nb.get("metadata", {})
lang = (
    meta.get("language_info", {}).get("name")   # set by kernel at runtime
    or meta.get("kernelspec", {}).get("language") # set at notebook creation
    or "code"                                     # fallback
)

A standard Python notebook produces ```python blocks; the "code" fallback applies only when metadata is absent entirely.