Improve numbering numId lookup

[Sean-Kenneth-Doherty]

Fixes #940.

CT_Numbering._next_numId gathered existing IDs in a list, then checked membership for every candidate ID. When numbering definitions grow large and contiguous, that makes the lookup quadratic in the number of <w:num> elements.

This changes the collection to a set so membership is constant-time while preserving the existing behavior of returning the first gap, starting at 1. I also added focused oxml tests for empty, contiguous, and gapped numbering definitions.

Local validation:

• uv run --with pytest --with ruff --with-editable . python -m pytest tests/oxml/test_numbering.py -q
• uv run --with pytest --with ruff --with-editable . python -m pytest tests/oxml/test_numbering.py tests/parts/test_numbering.py -q
• uv run --with pytest --with ruff --with-editable . python -m ruff check src/docx/oxml/numbering.py tests/oxml/test_numbering.py
• git diff --check

Local timing sanity check on 5,000 contiguous IDs, isolating the membership loop:

• list membership path: 0.387197s
• set membership path: 0.008164s

[Sean-Kenneth-Doherty]

Moved out of draft after local validation:

• uv run --with pytest --with ruff --with-editable . python -m pytest tests/oxml/test_numbering.py tests/parts/test_numbering.py -q -> 8 passed
• uv run --with pytest --with ruff --with-editable . python -m pytest tests/oxml/test_numbering.py -q -> 3 passed
• uv run --with pytest --with ruff --with-editable . python -m ruff check src/docx/oxml/numbering.py tests/oxml/test_numbering.py
• git diff --check

Timing sanity check on 5,000 contiguous IDs, isolating the membership loop:

• list membership path: 0.797563s
• set membership path: 0.003804s
• speedup: 209.7x