Stack dependency

This draft is stacked on #4408 and should be reviewed and merged after that PR. Once #4408 lands, this branch will be rebased onto main so the final diff contains only the parser-free POS change.

Summary

Skip spaCy's dependency parser in pos_tag() while retaining the tok2vec and tagger components that produce token.tag_.

The dependency parser runs after the tagger, and pos_tag() immediately discards its output. Disabling it removes unnecessary statistical work without changing returned POS tags.

Performance

Benchmarked against the tokenizer-only parent branch on an 8-vCPU AMD EPYC host with 15 GiB RAM and Python 3.12. The spaCy model was warmed, NLP result caches were cleared before every iteration, and each full-matrix result is the median of three runs.

A seven-iteration follow-up on the short workloads measured HTML 1.5% faster and PPTX 1.6% slower, confirming both are effectively neutral. The repeatable gains are in the text-heavy TXT and DOCX paths.

Compared with upstream/main, the combined tokenizer-only and parser-free-POS stack improves the representative aggregate from 85.801s to 70.998s: approximately 17.3% faster / 1.21x.

Output parity was exact for all four documents.

Testing

• 128 focused NLP, text-classification, and benchmark tests pass.
• POS tags are compared with and without the dependency parser across six representative inputs.
• A regression test verifies that pos_tag() disables the parser.
• Existing tokenizer parity, truncation, caching, and benchmark tests continue to pass.
• Ruff lint and formatting checks pass.