Add inference-time sequence packing support

[vivekkalyanarangan30]

Summary

• Add a new inference-packing utility that builds packed batches with block-diagonal masks and helpers to unpack outputs.
• Expose configuration knobs on the GLiNER API so packing can be toggled globally or per-call.
• Wire the encoder to use packed execution when configured, including automatic pack/unpack around the transformer forward pass.
• Results are identical with or without packing (verified by tests).

Benchmarks

All runs on CPU, roberta-base, batch_size=64, max_length=512.

Scenario	Baseline tokens/s	Packed tokens/s	Speedup	Padding ↓
short_zipf	2.00e+03	3.88e+03	1.94×	61.5% → 12.0%
short_uniform	2.44e+03	3.47e+03	1.42×	45.9% → 13.4%
mixed_tail	6.18e+02	3.43e+03	5.55×	87.4% → 19.6%
flat_long	4.94e+03	4.10e+03	0.83×	0.0% → 0.0%

👉 Packing yields 1.4–5.5× throughput improvements when input lengths are short or skewed, while performance is neutral (or slightly worse) when all sequences are long and uniform.

• document how to enable and benchmark inference-time sequence packing in README_Extended
• bench/bench_gliner_e2e.py to support full GLiNER models benchmarking

---

[vivekkalyanarangan30]

@urchade hopefully this helps.

[urchade]

@Ingvarstep 👀

[Ingvarstep]

@vivekkalyanarangan30 , awesome job, thanks for contributing!