Improve Urdu language data and add a community training example

[mwzkhalil]

Summary

• Strengthen spacy.blank(ur) with copied tokenizer tables, Latin abbreviations, unspaced Urdu punctuation (including after harakat), and a like_num policy that treats U+060C as a list comma rather than a thousands separator.
• Keep digit detection on Python str.isdigit for ASCII, Arabic-Indic, and Eastern Arabic-Indic digits; only واں/ویں forms with a known numeric stem count as numbers.
• Add examples/urdu, a community training project with JSONL/CoNLL-U/BIO converters, official-split preservation, a CPU overfit diagnostic, and an optional XLM-R config. Licensed corpora and trained weights are not included.

This is language data and an example project, not an official Explosion Urdu pipeline.

Test plan

• python -m pytest spacy/tests/lang/ur examples/urdu/tests -q
• Confirm spacy.blank(ur) splits لفظ۔لفظ and ہےً۔کیا, keeps Dr. intact, and does not treat 1،2 or کتابواں as a single number
• From examples/urdu, run python -m spacy project run overfit and treat the score as training-set learnability only
• Confirm generated corpus/, training/, and *.spacy artefacts remain untracked