feat: preserve markdown tables and add raw ProseMirror ingest flags (#42)

[teslakoile]

Summary

Fixes #42. Markdown tables sent via huly documents describe --set /
--set-file now parse into real ProseMirror table / tableRow /
tableHeader / tableCell nodes instead of being stored as paragraphs
full of literal | pipes. A raw-ingest escape hatch is also added for
ProseMirror features the markdown parser does not cover.

Changes

Path B — raw ProseMirror ingest (shipped first)

• New flags on documents describe:
  • --set-raw JSON_STRING — pass arbitrary ProseMirror JSON straight to
    client.set_content / client.create_content.
  • --set-raw-file PATH — same, but read from disk.
• Mutual-exclusion guard now covers all four write flags (--set,
  --set-file, --set-raw, --set-raw-file).
• Minimum structural validation: the root must be
  {"type": "doc", "content": [...]}. Invalid JSON or wrong shape is
  rejected with a clear error before any network IO.
• Only src/huly_cli/commands/documents.py touched on the command side.
  client.py is untouched — set_content / create_content already
  accept arbitrary ProseMirror JSON.

Path A — GFM table parsing in markdown_to_prosemirror

• markdown_to_prosemirror in src/huly_cli/markup.py recognises GFM
  pipe tables:

  | a | b |
  | --- | --- |
  | 1 | 2 |
  

  First row → tableHeader, body rows → tableCell, alignment row
  consumed but not stored (1x1 cells per the issue non-goal).
• Cells are wrapped in a single paragraph so inline marks (bold,
  italic, links, code) survive.
• Escaped pipes (\|) inside cells are preserved as literal pipes.
• prosemirror_to_markdown already emits GFM tables via _table_to_md,
  so markdown → PM → markdown round-trips; a test asserts exact
  equality after .strip().
• README.md and skills/huly-cli/SKILL.md updated.

Test plan

• uv run pytest -x -q — 293 passing (up from 278: +15 new tests)
• uv run ruff format . && uv run ruff check . clean
• Parser-level tests in tests/test_markup.py:
  • basic table → table node
  • header/body cell type distinction
  • {"colspan": 1, "rowspan": 1} attrs
  • alignment row with :---/---: consumed
  • inline marks inside cells preserved
  • escaped-pipe (\|) cells
  • markdown → PM → markdown round-trip (exact equality after strip)
  • table followed by paragraph
  • non-table markdown unchanged (regression guard)
  • single | row with no separator stays a paragraph
  • | in plain prose stays a paragraph
  • uneven rows padded on round-trip
• Command-layer tests in tests/test_issue_write_path.py:
  • --set-raw writes verbatim (markup payload handed to create_content
    is the exact JSON, parsed result contains a table node)
  • --set-raw-file reads and writes
  • --set-raw-file with missing file exits 1
  • --set-raw with invalid JSON exits 1, no network call
  • --set-raw with non-doc root exits 1
  • --set-raw with a doc root missing content exits 1
  • all six mutual-exclusion pairs parameterised and asserted
  • existing --set "" --set-file regression updated to the new
    four-flag error wording
  • integration: --set with a markdown table emits a PM table node
    with tableHeader header row and tableCell body row

Closes #42